A Review of AI-Driven Engineering Modelling and Optimization: Methodologies, Applications and Future Directions

Abstract

Engineering is suffering a significant change driven by the integration of artificial intelligence (AI) into engineering optimization in design, analysis, and operational efficiency across numerous disciplines. This review synthesizes the current landscape of AI-driven optimization methodologies and their impacts on engineering applications. In the literature, several frameworks for AI-based engineering optimization have been identified: (1) machine learning models are trained as objective and constraint functions for optimization problems; (2) machine learning techniques are used to improve the efficiency of optimization algorithms; (3) neural networks approximate complex simulation models such as finite element analysis (FEA) and computational fluid dynamics (CFD) and this makes it possible to optimize complex engineering systems; and (4) machine learning predicts design parameters/initial solutions that are subsequently optimized. Fundamental AI technologies, such as artificial neural networks and deep learning, are examined in this paper, along with commonly used AI-assisted optimization strategies. Representative applications of AI-driven engineering optimization have been surveyed in this paper across multiple fields, including mechanical and aerospace engineering, civil engineering, electrical and computer engineering, chemical and materials engineering, energy and management. These studies demonstrate how AI enables significant improvements in computational modelling, predictive analytics, and generative design while effectively handling complex multi-objective constraints. Despite these advancements, challenges remain in areas such as data quality, model interpretability, and computational cost, particularly in real-time environments. Through a systematic analysis of recent case studies and emerging trends, this paper provides a critical assessment of the state of the art and identifies promising research directions, including physics-informed neural networks, digital twins, and human–AI collaborative optimization frameworks. The findings highlight AI’s potential to redefine engineering optimization paradigms, while emphasizing the need for robust, scalable, and ethically aligned implementations.

1. Introduction

Engineering optimization is undergoing a profound transformation driven by the rapid advancement and integration of artificial intelligence technologies [1,2]. Traditional engineering optimization approaches, such as gradient-based algorithms, often limited by computational complexity, simplified assumptions, high-dimensional problems [3] and disciplinary silos, are increasingly being augmented, and in some cases replaced by data-driven methodologies that can identify complex patterns and relationships beyond human analytical capabilities. This paradigm shift represents a fundamental change in how engineers approach design, simulation, and optimization across virtually all engineering disciplines [4].

The emergence of AI as a transformative force in engineering optimization coincides with growing complexity in engineering systems and increasing demands for efficiency, sustainability, and customization [5]. Classical optimization techniques struggle with high-dimensional, non-linear, and multi-physics problems, whereas AI algorithms (e.g., Bayesian optimization, deep learning models, generative AI and large-language models) demonstrate remarkable capabilities in navigating complex design spaces and identifying non-intuitive solutions [6]. The integration of Industry 4.0 technologies, such as IoT sensors, digital twins, and edge computing, has further accelerated this transition, generating vast datasets that fuel AI algorithms while providing implementation frameworks for optimized solutions [7].

This review paper provides a comprehensive overview of AI-driven optimization methodologies and their applications across engineering domains. Unlike previous reviews [6,8,9,10] focused on specific disciplines or isolated applications, this work adopts an interdisciplinary perspective, identifying common challenges, transferable methodologies, and emerging opportunities at the intersection of AI and engineering optimization. We summarized the frameworks on the integration of AI and engineering optimization and systematically analyzed the technical foundations of AI-driven optimization, surveyed cutting-edge applications across engineering disciplines, critically evaluated implementation challenges, and identified promising research directions. By synthesizing insights from diverse engineering fields, this review aims to bridge knowledge gaps between AI researchers and engineering practitioners while facilitating cross-disciplinary innovation in optimization methodologies.

Despite these advancements, challenges remain in areas such as data quality, model interpretability, and computational cost, particularly in real-time environments [8,11] as well as environmental sustainability. Although recent case studies and emerging trends provide valuable insights, there is still a gap in the critical assessment of the state of the art. This review identifies several promising research directions, including physics-informed neural networks, digital twins, and human–AI collaborative optimization frameworks. Overall, the findings highlight AI’s potential to redefine engineering optimization paradigms while emphasizing the need for robust, scalable, and ethically aligned implementations.

The remainder of this paper is organized as follows. Section 2 introduces the primary AI techniques employed in engineering optimization. Section 3 summarizes the existing frameworks for integrating artificial intelligence (AI) with engineering optimization, while Section 4 briefly presents the optimization algorithms that are commonly combined with these AI methods. Section 5 reviews applications of AI-driven optimization across various engineering domains. Section 6 discusses key challenges and limitations of AI-driven engineering optimization, and Section 7 outlines directions for future research. Finally, Section 8 concludes the paper and highlights potential avenues for continued investigation.

2. Overview of AI Modelling Techniques

2.1. Introduction

The field of AI (see Figure 1) has obtained increasing attention due to its ability to efficiently analyze and to act upon the vast volumes of data generated [12]. AI has been the subject of research since the 1950s, but the recent surge in interest is largely driven by advances in the subfield of ML and LLMs, as well as supporting factors such as enhanced data storage capabilities and computational power.

Bagheri et al. summarized the classification of AI and ML techniques used to improve membrane-fouling control [13], as the related to optimization is summarized in Figure 2. Essentially, training a machine learning model is an optimization problem in which the goal of training a model is to determine the optimal set of model parameters (weights and biases) that minimize a loss function, which measures prediction error, solved by using various optimization algorithms or techniques. However, the goals of machine learning and optimization differ. Machine learning focuses on developing predictive models, whereas optimization aims to determine the best parameters or design solutions for complex engineering problems. Therefore, although optimization plays a critical role in model training, machine learning and optimization are considered distinct fields due to their different objectives. In this section, we briefly introduce several key machine learning techniques that have been integrated into engineering optimization.

2.2. Artificial Neural Networks

Artificial neural networks (ANNs), also called neural networks (NNs), are mathematical models inspired by biological neural systems. They can be interpreted as mathematical functions that map an input space to an output space [14], as illustrated in Figure 3. A neural network typically consists of three types of layers. The first layer, the input layer, contains input neurons that receive data and pass it to the next layer. The hidden layer performs computations on the input data and forwards the results to the final layer. The output layer produces the network’s prediction or result. For advanced neural networks, such as deep learning, there are multiple hidden layers.

Each neuron can be viewed as a small computational unit that receives inputs, processes them, and generates an output (shown as Figure 4). The computation within a neuron involves weights, an activation function, and a cost function. One of the most used activation (transfer) functions is the logistic (sigmoid) function, given by:

$$output={\displaystyle \frac{1}{1+e^{-\left(\sum _i{\omega }_i{x}_i+{\omega }_0\right)}}},$$

(1)

where i is the index of the input, $x_i$ is the value of the i-th input, ${\omega }_i$ is the weight associated with that input, and ${\omega }_0$ represents the neuron’s bias.

Machine learning (ML) is used to determine optimal weight parameters for surrogate models that predict system behavior, performance metrics, and failure modes. These surrogate models can then serve as objective functions or constraint functions in engineering optimization problems [15]. For instance, in aerodynamic design, ML models can predict lift coefficients with remarkable accuracy, achieving a lower mean squared error (MSE) of 0.0002 and a higher R² of 0.9916 while reducing computational requirements by orders of magnitude compared to traditional computational fluid dynamics simulations (MSE of about 0.0031 and R² of 0.8516) [16].

Machine learning techniques are broadly categorized into supervised learning, unsupervised learning, and reinforcement learning, with additional important approaches including semi-supervised learning and deep learning.

• Supervised learning: Supervised learning is a type of machine learning in which the model is trained using labeled data—data that includes both inputs and their corresponding outputs [17]. The process of applying supervised learning to a real-world problem is illustrated in Figure 5. First, data collected from a specific problem is processed and organized into labeled datasets. These datasets are then divided into two parts: one for training and the other for validation. During training, the model’s parameters (such as weights and biases) are adjusted by selecting an appropriate training method. The trained model is subsequently evaluated and validated using the validation dataset. According to the objectives of the algorithms, supervised learning algorithms can be classified into two categories: classification and regression (see Figure 6) [18]. Regression-based supervised learning algorithms are commonly used in engineering optimization. Typical methods include linear and polynomial regression, support vector regression, neural networks, and others.

• Unsupervised learning: The objective of unsupervised learning algorithms is fundamentally different from that of supervised learning. Their primary aim is to explore the structure, patterns, or relationships within the data without prior knowledge of the “correct” answers [19].
• Reinforcement learning: The aims of reinforcement learning (RL) are distinct from both supervised and unsupervised learning. In RL, an agent learns to make decisions by interacting with an environment using a trial-and-error approach, gradually developing an optimal behavioral strategy based on the reward signals received from previous interactions [20].

2.3. Deep Neural Networks

Deep learning architectures/Deep Neural Networks (DNNs) have significantly expanded the scope and capabilities of engineering optimization, particularly due to their ability to process high-dimensional and unstructured data [21]. In this section, several DNN architectures will be discussed.

Convolutional Neural Networks (CNNs), introduced by LeCun (1989) in 1989 [22], have been proven exceptionally effective in spatial data analysis, enabling breakthroughs in fields ranging from fluid dynamics to materials science [23]. CNNs were originally developed for image processing, and a typical structure is shown in Figure 7. Figure 7 depicts a convolutional neural network (CNN) as a multi-layered neural network architecture, composed primarily of convolutional layers (layers) and sub-sampling layers, which work together to extract and down sample features from the input data.

Research has demonstrated that CNNs can not only predict engineering parameters but also internalize fundamental physical principles, such as the Kutta condition in aerodynamics, effectively learning the underlying physics governing system behavior. Jiang et al. pointed out that trained deep neural networks can serve as high-speed surrogate solvers [24].

Physics-Informed Neural Networks (PINNs) represent a significant advancement by embedding physical laws directly into the learning process, ensuring that model predictions adhere to fundamental constraints described by partial differential equations and guiding the optimization, architecture design and implementation of DNNs [25]. Even though there has been huge progress in computational fluid dynamics (CFD) in solving Navier–Stokes equations (NSEs) by using finite elements, spectral and meshless methods in the last 50 years, it is still very challenging to tackle high-dimensional problems governed by parametrized NSE [3]. Cai et al. demonstrated that PINNs are very efficient for inverse problems of three-dimensional fluid dynamics [3]. Huang and Wang (2023) summarized the applications of PINNs in power systems [26]. Laubscher (2021) created PINN models for simulating multi-species flow and heat transfer with the combination of different hyperparameter settings to find the best performing configuration [27].

2.4. Large Language Models

Large Language Models are trained on massive text datasets and have been widely demonstrated to have extraordinary capability in understanding, generating and translating human languages [28]. Figure 8 illustrates the working process of an LLM. User input texts are processed to clean the data and remove noise. After they are tokenized and used as inputs for a pre-training of the LLM. The output generated by the LLM is then translated into human languages. Chien et al. applied an LLM to generate initial parameters for an engineering design [29].

Table 1. Comparison of AI-modelling techniques.

Model	Full Name	Key Characteristics	Typical Inputs	Strengths	Common Applications
NN	Neural Network	General term for models composed of interconnected neurons	Numerical, categorical	Flexible function approximation	Regression, classification
ANN	Artificial Neural Network	Fully connected feedforward networks	Structured numerical data	Simple, effective for low- to medium-dimensional problems	Regression
CNN	Convolutional Neural Network	Uses convolutional layers to capture spatial patterns	Images, grids, spatial data	Translation invariance, parameter efficiency	Image recognition, computer vision
PINN	Physics-Informed Neural Network	Embeds physical laws into the loss function	Spatial–temporal coordinates, boundary conditions	Data-efficient, physically consistent	Scientific computing, inverse problems, engineering simulations
LMM	Large Language Model	Very large transformer-based models trained on text	Natural language, code	Strong reasoning and generative ability	Text generation, code synthesis, optimization guidance

provides a summary of the key characteristics, typical inputs, strengths, and common applications of the AI modeling techniques discussed in this paper.

3. Example Frameworks of AI-Driven Engineering Optimization

In this section, several existing frameworks that combine artificial intelligence techniques with optimization algorithms for engineering optimization will be summarized to illustrate the application of AI in engineering optimization.

Krzywanski et al. provided a comprehensive review of recent advancements for computational methods for modelling, simulation and optimization of complex systems [30]. Their key findings highlight that integrating artificial intelligence with traditional computation methods opens several promising research directions, as shown in Figure 9, for complex engineering systems. These studies have demonstrated significant improvements in both accuracy and efficiency.

The common AI/ML-enabled framework for engineering optimization is similar to the model presented by Ibn-Mohammed et al. [31], as shown in Figure 10. In those frameworks, machine learning techniques are used to train models that serve as objective functions or constraints in engineering optimization problems [32].

For example, Jin et al. presented a data-driven evolutionary optimization framework, shown in Figure 11, which includes data collection, machine learning, and optimization [33]. Machine Learning techniques can be used to train functions from the collected data, which then serve as objective or constraint functions to construct an appropriate optimization model. This model can be solved using optimization techniques such as gradient-based algorithms or population-based algorithms, including evolutionary algorithms.

Notably, Jin et al. also emphasized that data or domain knowledge can be applied to nearly every component of an evolutionary algorithm (EA) to accelerate the search process [33].

Another example is illustrated in Figure 12. Nadian et al. developed an intelligent fuzzy-machine vision control system for a hybrid hot air–infrared dryer [34]. In this framework, two artificial neural networks (ANNs) were employed. The first ANN was responsible for predicting material characteristics, while the second predicted the sample’s moisture ratio. All prediction models were integrated into a genetic algorithm as fitness functions to optimize the control parameters of a fuzzy logic control system, aiming to minimize energy consumption.

Some commercial software packages, such as ANSYS, Abaqus (Simulia), and MSC Nastran, are widely used to calculate stress in engineering design. Chien et al. introduced a particularly interesting framework exploring the potential application of large language models (LLMs) for generating geometric parameter suggestions during the early stages of structural design [35]. Their proposed workflow, illustrated in Figure 13, integrates design recommendations produced by the LLM with validation via a finite element solver. The results demonstrated that, while a regression-based model required 252 samples to construct a reliable polynomial model, the LLM achieved near-optimal design suggestions using only 18 initial reference samples.

Hsu et al. proposed a framework, illustrated in Figure 14, that integrates artificial intelligence with finite element methods to optimize the properties of woven composites [35]. ABAQUS simulations, a finite element approach, were employed to generate the initial dataset. A dual-input CNN model was then developed to capture both material parameters and performance characteristics, while a Deep Q-Network (DQN) was used to enable adaptive decision-making.

From the above discussion, researchers have explored several ways to integrate machine learning into engineering optimization, shown in

Table 2. Existing Framework of AI-Driver Engineering Optimization.

No	Framework	Description
#1	AI-based Modelling	Machine learning models are used as objective and/or constraint functions in optimization problems
#2	AI-improved optimization	Machine learning techniques are employed to develop AI-enhanced optimization algorithms
#3	AI-based Model to Approximate complex engineering simulations	Machine learning is used to approximate computationally expensive engineering simulations, such as finite element analysis (FEA) and computational fluid dynamics (CFD), making the optimization of complex engineering problems feasible
#4	AI searches an initial solution	Machine learning techniques are employed to predict initial solutions or design parameters, which can significantly speed up the optimization process

. (1) A common approach is to train machine learning models that serve as surrogate objective or constraint functions within an optimization problem. (2) Machine learning can also be used to develop AI-enhanced optimization algorithms that improve search efficiency. (3) Many engineering simulations, such as finite element analysis (FEA) and computational fluid dynamics (CFD), are computationally expensive and therefore unsuitable for direct use within iterative optimization loops. Machine learning offers a solution by enabling the development of fast surrogate models that approximate these simulations. (4) Additionally, machine learning techniques, including large language models, can generate design parameters or initial solutions that accelerate the overall optimization process.

4. Optimization Techniques

4.1. Introduction

AI-Driven Optimization Techniques refer to optimization methods that leverage artificial intelligence (AI) algorithms to find optimal solutions for complex engineering, scientific, or industrial problems. Unlike traditional optimization methods that rely on explicit mathematical formulations, AI-driven approaches can handle nonlinear, high-dimensional, and computationally expensive problems, often using data-driven models.

Common AI-driven optimization techniques are metaheuristics (Figure 15), which can be categorized into two main groups [36]:

• Neighbourhood-based algorithms: These are local search methods that utilize the concept of a neighbourhood by exploring and tracking nearby candidate solutions. Examples include Simulated Annealing and Tabu Search.
• Population-based algorithms: These are nature-inspired methods that operate on a set of candidate solutions (a population). They can be further classified into evolutionary algorithms/computational methods and swarm intelligence.

Evolutionary algorithms (EAs), inspired by the principles of biological evolution, are a class of population-based metaheuristic search algorithms that provide powerful global optimization capabilities. They are particularly well-suited for complex, non-convex engineering problems with multiple local optima [37]. The family of EAs has grown to include many variants, such as genetic algorithms [38], evolutionary programming [39], evolutionary strategies, genetic programming [40], and differential evolution [36].

Swarm Intelligence (SI) techniques are inspired by the collective and cooperative behavior observed in nature [41]. Examples include Particle Swarm Optimization and Ant Colony Optimization.

Due to space limitations, only selected methods commonly applied to engineering optimization problems in combination with artificial intelligence are summarized here.

4.2. Genetic Algorithms

Genetic algorithms (GAs) have demonstrated remarkable effectiveness in multi-objective optimization scenarios where traditional gradient-based methods often struggle [42]. The process of a classic genetic algorithm is illustrated in Algorithm 1. GAs use a variety of operators during the search process, including crossover, mutation, and selection. In a genetic algorithm, the objective function of an optimization problem is referred to as the fitness function, and the decision variables are represented by chromosomes.

○

Crossover operators generate offspring by combining the genetic information of two parents randomly selected from the population.

○

Mutation operators help maintain the diversity of the population, preventing premature convergence.

○

Selection operators choose individuals to form a new population based on their fitness values, with individuals of higher fitness generally having a higher probability of survival in the next generation.

Algorithm 1: Genetic Algorithm

Inputs:$\mathrm{Population}\mathrm{Size}\left(n\right),\mathrm{Maximum}\mathrm{number}\mathrm{of}\mathrm{iterations}(T_{max}$)    $\mathrm{Crossover}\mathrm{probability}(p_c$$),\mathrm{Mutation}\mathrm{probability}(p_m)$ Output:$\mathrm{Global}\mathrm{solution},{\overrightarrow{\mathit{x}}}^g$

1.  Begin 2.   $\mathrm{Generate}\mathrm{initial}\mathrm{population}\mathrm{of}\mathrm{n}\mathrm{Chromosomes},{\overrightarrow{x}}_i(i=\mathrm{1,2},\cdots ,n)$ 3.   Set iteration counter t = 0 4.   $\mathrm{Compute}\mathrm{the}\mathrm{fitness}\mathrm{value}{F}_i$ of each chromosome 5.   while (t < $T_{max}$) 6.     Select n chromosomes from the population based fitness 7.     Apply crossover operation on random selected pair with crossover probability 8.     Apply mutation operation on random selected chromosome with mutation probability 9.     t = t + 1; 10.   end while 11.   $\mathrm{Search}\mathrm{the}\mathrm{global}\mathrm{solution}{\overrightarrow{\mathit{x}}}^g$ from the population 12.   $\mathrm{Return}\mathrm{the}\mathrm{best}\mathrm{solution}{\overrightarrow{\mathit{x}}}^g$ 13. end

4.3. Particle Swarm Optimization

Particle Swarm Optimization (PSO), a prominent algorithm within the field of swarm intelligence, was first introduced by Eberhart and Kennedy in 1995. It is a population-based stochastic optimization method inspired by the social behaviors observed in flocks of birds [43,44,45].

In the standard PSO algorithm, the population consists of multiple particles. Each particle retains personal information, including its position, movement velocity, and a memory of its best solution found so far. The velocity of each particle is updated at every iteration using the following equation.

$${\overrightarrow{v}}_i\left(t+1\right)={\omega \overrightarrow{v}}_i\left(t\right)+c_1{r}_1\left({\overrightarrow{x}}_i-{\overrightarrow{x}}_i^p\right)+c_1{r}_1\left({\overrightarrow{x}}_i-{\overrightarrow{x}}^g\right),$$

(2)

where ${\overrightarrow{x}}_i$ represents the position of the i-th particle and ${\omega \overrightarrow{v}}_i\left(t\right)$ is its velocity at iteration t. ${\overrightarrow{x}}_i^p$ denotes the personal best position achieved by the i-th particle, while ${\overrightarrow{x}}^g$ represents the global best position found by the entire swarm. The parameter $\omega$ is the inertia weight, which controls the influence of the particle’s previous velocity. The parameters $c_1$ and $c_2$ are acceleration coefficients that determine the relative impact of the personal and global best positions, respectively. $r_1$ and $r_2$ are independent random values uniformly distributed in the range [0, 1]. The pseudocode of the standard Particle Swarm Optimization algorithm is presented in Algorithm 2.

Algorithm 2: Particle Swarm Optimization

Inputs:$\mathrm{Population}\mathrm{Size}\left(n\right),\mathrm{Maximum}\mathrm{number}\mathrm{of}\mathrm{iterations}(T_{max}$)     $\mathrm{Weight}\mathrm{factors}w,c_1,c_2$ Output:$\mathrm{Global}\mathrm{solution},{\overrightarrow{\mathit{x}}}^g$

1. Begin 2.   $\mathrm{Generate}\mathrm{initial}\mathrm{population}\mathrm{of}\mathrm{n}\mathrm{Chromosomes},{\overrightarrow{x}}_i(i=\mathrm{1,2},\cdots ,n)$ 3.   Set iteration counter t = 0 4.   $\mathrm{Compute}\mathrm{the}\mathrm{fitness}\mathrm{value}{F}_i$ of each chromosome 5.   $\mathrm{Set}\mathrm{personal}\mathrm{best}{{\overrightarrow{x}}_i^p=\overrightarrow{x}}_i$$(i=\mathrm{1,2},\cdots ,n)$ 6.   $\mathrm{Generate}\mathrm{random}\mathrm{velocities}{\overrightarrow{v}}_i(i=\mathrm{1,2},\cdots ,n)$ 7.   $\mathrm{Find}\mathrm{the}\mathrm{global}\mathrm{best}{\overrightarrow{x}}_i^g$ from the population 8.   while (t < $T_{max}$) 9.     for i = 1: n do 10.      $\mathrm{Generate}\mathrm{random}\mathrm{numbers},r_1,r_2\in \left[\mathrm{0,1}\right]$ 11.      $\mathrm{Update}\mathrm{velocity}{\overrightarrow{v}}_i\left(t+1\right)=w\times {\overrightarrow{v}}_i+c_1\times r_1\left({\overrightarrow{x}}_i-{\overrightarrow{x}}_i^p\right)+c_2\times r_2({\overrightarrow{x}}_i-{\overrightarrow{x}}_i^g$) 12.      $\mathrm{Update}\mathrm{position}{\overrightarrow{x}}_i={\overrightarrow{x}}_i+{\overrightarrow{v}}_i\left(t+1\right)$ 13.      $\mathbf{if}{\left(F\right(\overrightarrow{x}}_i)>F({\overrightarrow{x}}_i^p)$ then 14.         ${\overrightarrow{x}}_i^p={\overrightarrow{x}}_i$ 15.      end if 16.      $\mathbf{if}\left({F(\overrightarrow{x}}_i\right)>F\left({\overrightarrow{x}}_i^g\right)$ ) then 17.         ${\overrightarrow{x}}_i^g={\overrightarrow{x}}_i$ 18.      end if 19.     end for 20.     t = t + 1; 21.   end while 22.   $\mathrm{Search}\mathrm{the}\mathrm{global}\mathrm{solution}{\overrightarrow{\mathit{x}}}^g$ from the population 23.   $\mathrm{Return}\mathrm{the}\mathrm{best}\mathrm{solution}{\overrightarrow{\mathit{x}}}^g$ 24. end

4.4. Differential Evolution

Differential Evolution (DE) was originally developed by Storn and Price in 1995 and later published in 1997 [46] for continuous minimization problems. The general procedure of DE is like that of a Genetic Algorithm, as it also involves mutation, crossover, and selection [36].

• Mutation: For each target individual in the population, a mutant vector is generated by adding the scaled difference between two randomly selected individuals to a third individual. A common mutation formula is:

$${\overrightarrow{v}}_i={\overrightarrow{x}}_{r1}+F·\left({\overrightarrow{x}}_{r2}-{\overrightarrow{x}}_{r3}\right)$$

(3)

where ${\overrightarrow{x}}_{r1}$, ${\overrightarrow{x}}_{r2}$, and ${\overrightarrow{x}}_{r3}$ are distinct random individuals, and F is the scaling factor.

• Crossover: A trial vector is produced by combining components of the mutant vector and the target vector, typically using binomial or exponential crossover.

$$x_{i,j}=\left\{\begin{array}{cc}{v}_{i,j}& if{rand}_j\le C_r\\ x_{i,j}& otherwise\end{array}\right.$$

(4)

• Selection: The trial vector is compared to the target vector. If the trial vector yields a better (lower) objective function value, it replaces the target vector in the next generation; otherwise, the original target vector is retained.

$${\overrightarrow{x}}_i=\left\{\begin{array}{cc}{\overrightarrow{v}}_i& iff\left({\overrightarrow{v}}_i\right)\le f\left({\overrightarrow{x}}_i\right)\\ {\overrightarrow{x}}_i& otherwise\end{array}\right.$$

(5)

4.5. Ant Colony Optimization

In the early 1990s, M. Dorigo and colleagues introduced ant colony optimization (ACO) [47], a nature-inspired metaheuristic novel that mimics how real ants find the shortest path between their nest and a food source, to solve challenging combinatorial optimization (CO) problems. Algorithm 3 presents the ACO metaheuristic. After initialization, each iteration consists of three phases: solution construction by the ants, optional improvement via local search, and pheromone update [48].

Algorithm 3: Ant Colony Optimization

Inputs:$\mathrm{Problem}\mathrm{Graph}G=(V,E)$$;\mathrm{Number}\mathrm{of}\mathrm{ants}\left(\mathrm{m}\right);\mathrm{Pheromone}\mathrm{evaporation}\mathrm{rate}(\rho \in (\mathrm{0,1})$);    $\mathrm{Pheromone}\mathrm{importance}\left(\alpha \right)$$;\mathrm{Heuristic}\mathrm{importance}\left(\beta \right)$$;\mathrm{Maximum}\mathrm{number}\mathrm{of}\mathrm{iterations}\left(T_{max}\right)$ Output:$\mathrm{Global}\mathrm{solution},{\overrightarrow{\mathit{x}}}^g$

1. Begin 2.   Initialize pheromone trails 3.   Set iteration counter t = 0 4.   while (t < Max) 5.     Construct ant solutions 6.     Apply local search 7.     Update pheromones 8.     t = t + 1; 9.   end while 10.  $\mathrm{Search}\mathrm{the}\mathrm{global}\mathrm{solution}{\overrightarrow{\mathit{x}}}^g$ from the population 11.  $\mathrm{Return}\mathrm{the}\mathrm{best}\mathrm{solution}{\overrightarrow{\mathit{x}}}^g$ 12. end

4.6. Multi-Objective Evolutionary Algorithms

Multi-Objective Evolutionary Algorithms (MOEAs) are a class of optimization algorithms designed to solve problems with multiple conflicting objectives simultaneously by finding or approximating the Pareto front, which represents the trade-offs among objectives [49]. A solution is considered Pareto-optimal if no other solution can improve one objective without worsening at least one other. Among the most popular MOEAs are:

• NSGA-II (Non-dominated Sorting Genetic Algorithm II) [50]: A widely used MOEA that employs fast non-dominated sorting, elitism, and crowding distance to efficiently approximate a well-distributed Pareto front.
• SPEA2 (Strength Pareto Evolutionary Algorithm 2) [51]: An improved version of the original SPEA [52] which uses an external archive, refined fitness assignment based on Pareto dominance, and density estimation to maintain both convergence and diversity on the Pareto front.
• MOEA/D (Multi-objective Evolutionary Algorithm based on Decomposition). MOEA/D decomposes a multi-objective problem into multiple scalar optimization subproblems and solves them simultaneously, promoting both convergence and diversity across the Pareto front [53].

4.7. Reinforcement-Learning-Based Algorithm for Optimization

Reinforcement learning is a machine learning method in which an agent learns to make decisions by interacting with an environment to achieve a specific goal [54]. As shown in Figure 16, once the agent takes an action, the environment responds by providing a new state and a reward (either positive or negative). The agent then updates its strategy, known as a policy, in order to maximize future rewards.

Based on the reinforcement learning framework shown in Figure 16, Yonekura and Hattori developed a reinforcement learning-based optimization framework [55], illustrated in Algorithms 4 and 5. This framework consists of two phases: the training phase and the searching phase. In the algorithms, x represents the decision variables or parameter vector, f is the objective function, and s denotes the state. The training algorithm is designed to train a deep Q-network, while the searching algorithm is used to maximize the objective function values.

Algorithm 4: Training Phase

1. while (true) 2.   ${\overrightarrow{x}}_0\leftarrow$$\mathrm{initial}\mathrm{vector},f_0,s_0\leftarrow G\left({\overrightarrow{x}}_0\right)$ 3.   $\mathrm{for}i=\mathrm{1,2},\cdots ,n)$ do 4.     $\mathrm{Choose}{\overrightarrow{\alpha }}_i$$\mathrm{based}\mathrm{on}Q(s_{i-1},{\overrightarrow{x}}_{i-1})$ 5.     ${\overrightarrow{x}}_i\leftarrow {\overrightarrow{x}}_{i-1}+{\overrightarrow{\alpha }}_i,f_i,s_i\leftarrow G\left({\overrightarrow{x}}_i\right)$    ${\overrightarrow{r}}_i=H(s_i,s_{i-1},\cdots ,f_i,f{s}_{i-1},\cdots )$ 6.     Update Q functions 7.   end for 8.   if termination criterion is satisfied then 9.     return Q 10.   end if 11. end while

Algorithm 5: Search Phase

1.   ${\overrightarrow{x}}_0\leftarrow$$\mathrm{initial}\mathrm{vector},f_0,s_0\leftarrow G\left({\overrightarrow{x}}_0\right)$ 2.   $\mathrm{for}i=\mathrm{1,2},\cdots ,n)$ do 3.     $\mathrm{Choose}{\overrightarrow{\alpha }}_i$$\mathrm{based}\mathrm{on}Q(s_{i-1},{\overrightarrow{x}}_{i-1})$ 4.     ${\overrightarrow{x}}_i\leftarrow {\overrightarrow{x}}_{i-1}+{\overrightarrow{\alpha }}_i,f_i,s_i\leftarrow G\left({\overrightarrow{x}}_i\right)$ 5.   end for 6.   if termination criterion is satisfied then 7.     $\mathbf{return}{\overrightarrow{x}}_i$ 8.   end if

4.8. Summary of Optimization Techniques

Table 3. Comparison of Optimization techniques.

Algorithms	Key Characteristics	Parameters	Strengths	Common Applications
Genetic Algorithm (GA)	Evolutionary algorithm based on natural selection and genetics	Population Size, mutation rate, crossover rate	Good global search, flexible, widely used	Optimization problems, engineering design, scheduling
Particle Swarm Optimization (PSO)	Swarm intelligence inspired by social behavior of birds/fish	Population Size, weight factors for inertial position and global position.	Fast convergence, simple to implement	Continuous optimization, neural network training, control systems
Differential Evolution (DE)	Evolutionary algorithm using vector differences for mutation	Images, grids, spatial data	Robust, easy to implement, good for continuous problems	Parameter optimization, engineering design, machine learning
Ant Colony Optimization (ACO)	Swarm-based algorithm inspired by ant foraging behavior	Pheromone trails, heuristic info, number of ants, evaporation rate	Good for combinatorial optimization	Traveling Salesman Problem, routing, scheduling
Non-dominated Sorting Genetic Algorithm II (NSGAII)	Multi-objective GA with elitism and fast non-dominated sorting	Population Size, mutation rate, crossover rate	Efficient multi-objective optimization, maintains diversity	Multi-objective engineering optimization
Multi-objective Evolutionary Algorithm based on Decomposition (MOEA/D)	Decomposes multi-objective problem into scalar subproblems	Population size, weight vectors, neighborhood size	Good convergence and diversity balance, scalable	Multi-objective engineering optimization

provides a brief summary of the key characteristics, control parameters, strengths, and common applications of the optimization techniques discussed in this paper.

5. Engineering Applications of AI-Driven Optimization

5.1. Mechanical and Aerospace Engineering

In mechanical and aerospace engineering, AI-driven optimization has revolutionized traditional approaches to design, manufacturing, and maintenance [56]. Generative design algorithms leverage AI to autonomously generate and evaluate countless design alternatives based on specified constraints and performance requirements, dramatically accelerating the conceptual design phase.

5.1.1. Machine Learning as a Substitute for the Finite Element Method

Stress analysis is a fundamental aspect of mechanical system design. Finite Element Analysis (FEA) is commonly employed to evaluate stress in complex structures and systems, supporting design optimization, maintenance planning, and safety assessments across various industries, including aerospace, automotive, architecture, and biomedical engineering. Badarinath et al. proposed an FEA-based ML approach to estimate the distress distribution over the entire system [31]. Their surrogate FEA, based on machine learning algorithms, can accurately predict the response of the beam, with artificial neural networks yielding the highest accuracy. Hsu et al. proposed an FEA-based ML framework, shown in Figure 14, that integrates artificial intelligence with finite element methods to optimize the properties of woven composites [35]. Shah et al. applied machine learning to the design of pressure equipment and developed ML models trained on 605 data samples obtained from ANSYS Mechanical (FEA analysis) [57]. The results demonstrated that the FEA-based ML model can significantly accelerate the analysis process while providing accurate predictions of membrane stresses.

5.1.2. Geometric Configurations

Granados-Ortiz et al. developed a machine learning–aided design optimization (MLADO) model for the design of a mechanical micromixer [58]. A random forest classifier was trained to predict which geometric configurations would lead to vortex shedding. A multi-objective optimization problem was formulated to minimize the required pumping power and maximize the mixing efficiency under given design constraints, and it was solved using the NSGA-II algorithm. Their results showed that the optimization time was massively reduced, from days to less than one minute. Granados-Ortiz et al. further noted that their MLADO framework can be readily adapted for other similar mechanical design applications [58].

Topology optimization is widely used to generate algorithmically optimized structures in mechanical design. However, it often requires extensive manual setup and careful tuning of parameters to ensure proper algorithmic performance and convergence. To address this, Lynch et al. developed a machine learning–based framework that recommends optimal tuning parameters (such as setup or control parameters) to users, thereby reducing the costly trial-and-error typically involved in manual parameter selection [59].

5.1.3. Turbine Engine

Kosowski et al. proposed a general efficient system for designing turbine cascades and stages [60]. The design approach was based on evolutionary algorithms, as well as artificial neural networks, and shown to be efficient and computationally inexpensive compared with computational fluid dynamics calculations.

Xu et al. reviewed the potential of applying machine learning techniques to turbine cooling design [61]. The design of turbine blades involves multiple iterative optimization cycles. Machine learning can effectively utilize historical data to develop high-precision models, thereby significantly reducing the time required for performance evaluation and enhancing overall optimization efficiency.

Du et al. developed a dual convolutional neural network to predict turbine blade performance and applied a gradient-based optimization algorithm to optimize rotor blade designs [62]. Their results show that the optimization process can be completed within 38 s, and the efficiencies of the two optimized blades reach 89.29% and 88.92%, respectively, demonstrating the feasibility of the proposed method.

5.1.4. Heat Pump

Patel et al. focused on AI-enhanced optimization of heat pump sizing and design for application-specific requirements [63]. The study systematically combined advanced machine learning techniques with domain-specific heuristics to optimize heat pump configurations for performance, energy efficiency, and cost, achieving up to a 25% improvement in energy efficiency.

5.1.5. Unmanned Aerial Vehicles

In the literature, numerous studies have employed machine learning techniques to address design challenges in aerial vehicle development. Sharma and Hosder (2021) examined the feasibility of neural network models for predicting key airliner configuration parameters, including maximum take-off weight, fuselage length, thrust, and aspect ratio [64]. Oroumieh et al. proposed an approach integrating fuzzy logic and neural networks to determine wing area and engine thrust, which they validated through an application to a class of light business jets [65]. Boutemedjet et al. focused on small UAV design, deriving design parameters statistically from historical UAV data and optimizing the wing planform via a neural network-based aerodynamic model [66]. Similarly, Setayandeh developed a neural network-based meta-model of the multidisciplinary design and analysis module to evaluate and obtain the handling qualities of a small UAV [67]. Karali et al. integrated deep neural networks with a multi-objective genetic algorithm to optimize UAV configurations across various mission scenarios, including intelligence, surveillance, and reconnaissance [68].

The selected references in this paper related to mechanical and aerospace engineering, along with their key characteristics and performance metrics, are summarized in

Table 4. AI-Driven Engineering Optimization in Mechanical and Aerospace Engineering.

Reference	Key Characteristics	Problems	Performances
Badarinath et al. [31]	ML + FEA	One-Dimensional Beam	R2 is over 0.98
Hsu et al. [35]	CNN + reinforcement learning-based optimization	Optimization of woven composites	R2 is over 0.96; 2.37-fold increase in strain energy density; 267-fold acceleration in simulation time
Shah et al. [57]	ML + ANSYS (FEA)	Design of Pressure Equipment	R2 is over 0.99865
Granados-Ortiz et al. [58]	ML + NSGA-II	multi-objective optimization of mechanical micromixer	Efficiency improvement by 50%
Lynch et al. [59]	ML + Bayesian optimization	Topology Optimization	Reduce the total number of “wasted” tuning runs
Kosowski et al. [60]	ANN + evolutionary algorithms	Design of turbine cascades and stages	The optimization time was massively reduced, from days to less than one minute
Du et al. [62]	CNN + Gradient-based optimization	Rotor blade designs	The optimization time is within 38 s. R2 is over 0.99.
Patel et al. [63]	ML + Heuristics	Heat pump sizing and design	A 25% improvement in energy efficiency
Oroumieh et al. [65]	Fuzzy Logic + NN	Aircraft configuration design variables	Efffective to decrease aircraft design cycle time
Setayandeh [67]	NN + NSGA-II	UAV aerodynamic design	Reduce of computational costs by 94.1%
Karali et al. [68]	DNN + NSGA-II	UAV Design	R2 is 0.9971. The optimization process is 4–5 s.

.

5.2. Civil and Environmental Engineering

The civil and environmental engineering sectors have embraced AI-driven optimization to address challenges in infrastructure design, construction management, and environmental sustainability [69,70]. Concrete mix design is the process of determining the optimal proportions of cement, aggregates (sand and gravel), and water to achieve the desired strength, durability, and workability characteristics. Golafshani et al. developed a green mix design model to determine the constituents of rubbercrete using a machine learning–based ensemble model combined with constrained multi-objective grey wolf optimization algorithms, which outperformed the conventional M5P tree and MGEP models by 13.7% and 5.5%, respectively [71]. Zheng et al. proposed a multi-objective optimization model for concrete mix design that integrates machine learning with advanced techniques such as K-fold cross-validation, Bayesian hyperparameter optimization, regression feature elimination, and the C-TAEA algorithm [72]. Huang et al. proposed an artificial intelligence–based multi-objective optimization framework for the design of steel fiber reinforced concrete [73]. In their approach, support vector regression (SVR) was employed to predict both the compressive and flexural strengths, while the firefly algorithm was utilized to identify Pareto-optimal design solutions. Parhi et al. developed three machine learning models to optimize the critical parameters of self-compacting geopolymer concrete, with a primary focus on improving compressive strength [74]. The incorporation of supplementary materials in concrete presents a viable approach to reducing its environmental footprint. Saleh et al. employed machine learning (ML) algorithms to predict and formulate an empirical model for the compressive strength (CS) of concrete containing supplementary materials, aiming to optimize material strength [75].

Urban logistics play a pivotal role in the development of smart cities, aiming to enhance the efficiency and sustainability of goods delivery in urban environments. Mohsen proposed an innovative framework that integrates artificial intelligence (AI), autonomous vehicles (AVs), and Internet of Things (IoT) technologies to address these challenges [76]. The framework leverages real-time data from IoT-enabled infrastructure to optimize route planning, improve traffic signal control, and enable predictive demand management for delivery services.

In structural engineering, AI algorithms optimize material usage while maintaining structural integrity, leading to more sustainable and cost-effective designs. For construction materials, AI-driven approaches have been successfully applied to optimize concrete mixtures [77], with studies demonstrating the ability to predict compressive strength with high accuracy while reducing costs and environmental impact. These capabilities enable engineers to balance multiple competing objectives, including cost, durability, sustainability, and constructability.

In environmental engineering, AI optimization plays a crucial role in resource management and pollution control [78]. AI-driven wastewater treatment plants optimize process parameters in real-time to enhance treatment efficiency while reducing energy consumption [79]. Similarly, in water distribution systems, AI algorithms optimize pump schedules and control strategies to minimize energy usage while maintaining service levels [80]. These applications demonstrate how AI-driven optimization can simultaneously address economic, environmental, and operational objectives, contributing to more sustainable infrastructure systems.

The selected references in this paper related to Civil and Environmental Engineering, along with their key characteristics and performance metrics, are summarized in

Table 5. AI-Driven Engineering Optimization in Civil and Environmental Engineering.

Reference	Key Characteristics	Problems	Performances
Golafshani et al. [71]	ML + grey wolf optimization	Rubbercrete	Outperformed the conventional M5P tree and MGEP models by 13.7% and 5.5%, respectively
Zheng et al. [72]	ML + Bayesian optimization multi-objective optimization	Concrete mix design	R2 is over 0.98
Huang et al. [73]	SVR + Multi Objective Optimization + Firefly Algorithm	Steel fiber reinforced concrete	R2 is over 0.9142
Parhi et al. [74]	ML + Evolutionary Computations	Self-compacting geopolymer concrete
Mohsen [76]	AI + AV + IoT	Route planning
Kulkarni et al. [79]	DL	Wastewater Treatment Plants	85% accuracy

.

5.3. Electrical and Computer Engineering

The integration of AI-driven optimization in electrical and computer engineering has enabled significant advancements in energy systems, electronic design automation, and network management.

5.3.1. Blockchain

Blockchain is a distributed, decentralized digital ledger technology that records transactions across a network of computers in a secure, transparent, and tamper-resistant way, and Blockchain technology has gained significant attention with the introduction of cryptocurrency, Bitcoin, in 2006 [81]. Yuan et al. investigated the application of artificial intelligence technologies to address core challenges in blockchain systems—such as consensus mechanisms, algorithms, smart contracts, privacy protection, and data retrieval—particularly in relation to scalability, security, and privacy [82].

Proof-of-Stake (PoS) is a blockchain consensus mechanism that validates transactions and adds new blocks to the blockchain based on how many coins or tokens a participant holds and is willing to “stake” (lock up) as collateral. Arulkumaran et al. summarized the existing AI-driven framework for POS optimization: Dynamic validator assessment, performance optimization and risk management [83].

5.3.2. Semiconductor Design

The design phase accounts for approximately 42.7% of total semiconductor development resources [84]. According to comprehensive industry analyses, complex System-on-Chip (SoC) designs require an average of 81,340 engineer-hours, distributed across architectural planning (31.2%), implementation (38.5%), and verification (30.3%).

In electronic design automation, AI techniques have revolutionized circuit design and optimization [85]. Machine learning algorithms can predict circuit performance, optimize component placement and routing, and automate verification processes, dramatically reducing design cycles while improving performance.

5.3.3. Computer Science

Panwar examined the transformative potential of AI-driven query optimization in database management, highlighting how machine learning algorithms can intelligently predict and execute the most efficient query plans using both historical and real-time data [86].

Richardson et al. studied AI-driven optimization techniques to enhance software architecture design in complex systems, addressing scalability, flexibility, and performance while balancing competing objectives [87].

Cloud computing serves as a cornerstone of modern digital infrastructure. Chen et al. developed a collaborative scheduling algorithm for heterogenous cloud computing workflow that integrates deep learning (DL) to optimize workflow makespan, cost fairness and continuity in cloud computing while satisfing the constraints of task execution. Their results demonstrated the makespan was improved by 16.6% and the firness index was increased 5.3% [88].

The exponential growth in network complexity and data volume within modern digital ecosystems has heightened the need for AI-driven optimization techniques to enhance network performance and efficiency [89]. For example, by continuously monitoring network performance metrics and adapting configurations in real time, AI algorithms can optimize network topology to accommodate fluctuating demand and minimize congestion [90].

5.3.4. Electric Vehicles

The rapid growth in electric vehicle (EV) adoption necessitates advanced energy management strategies to ensure the sustainable, reliable, and efficient operation of charging infrastructure. Many metaheuristic techniques, such as Genetic Algorithms, Particle Swarm Optimization, and Ant Colony Optimization, have been shown to be effective in solving multi-objective EV charging problems, including peak load minimization and cost reduction, under constrained environments [91]. Sarker et al. proposed a hybrid AI-based framework to optimize residential EV charging systems by integrating Reinforcement Learning (RL), Linear Programming (LP), and real-time grid-aware scheduling [92]. Their results demonstrated a 31.5% reduction in peak transformer load, a decrease in voltage deviation from ±5.8% to ±2.3%, and an increase in solar utilization from 48% to 66%.

The selected references in this paper related to Electrical and Computer Engineering, along with their key characteristics and performance metrics, are summarized in

Table 6. AI-Driven Engineering Optimization in Electrical and Computer Engineering.

Reference	Key Characteristics	Problems	Performances
Chen et al. [88]	DL	Cloud Workflows	the makespan was improved by 16.6% and the firness index was increased 5.3%
Sarker et al. [92]	RL + LP + real-time grid-aware scheduling	Residential EV charging systems	31.5% reduction in peak transformer load, a decrease in voltage deviation from ±5.8% to ±2.3%, and an increase in solar utilization from 48% to 66%.

.

5.4. Chemical and Materials Engineering

Madika et al. highlighted that the development and application of artificial intelligence have significantly enhanced the design, modeling, and optimization of chemical and materials systems by accelerating discovery, reducing costs, and effectively addressing complex and nonlinear problems [93].

5.4.1. Material

Woven carbon fiber composites are increasingly utilized in advanced structural applications due to their exceptional strength-to-weight ratio and customizable design characteristics. Hsu et al. employed finite element methods to generate data elucidating the relationship between strain and stress and proposed a hybrid deep learning framework that integrates a dual-input Convolutional Neural Network (CNN) for predicting mechanical properties with a Deep Q-Network (DQN) for reinforcement learning-based optimization [35]. Their results demonstrated that the material’s strain energy density increased significantly, from 3590.78 J/m³ to 8527.85 J/m³, and the simulation time was reduced from 534 min to 2 min, a 267-fold speedup.

Kim et al. proposed a deep neural network–based forward design approach that enables efficient searching with a genetic algorithm to identify superior materials within a vast design space [94].

Magnesium (Mg) alloys show strong potential for lightweight structural and biomedical applications; however, their adoption is limited by issues such as poor corrosion resistance and complex deformation behavior. Wang et al. examined how Artificial Intelligence—specifically Machine Learning and Deep Learning—can accelerate the optimization of material properties by reducing experimental workload and speeding up the design of high-performance Mg alloys [95]. Despite these benefits, challenges remain in terms of data accessibility, model interpretability, and experimental validation.

The exceptional heat transfer properties of thermal metamaterials, derived from their sophisticated artificial structures, were the focus of a review by Zhu et al. [96]. The authors examined the significant potential of leveraging artificial intelligence and optimization algorithms to advance the design of these materials [96].

5.4.2. Chemical Process

AI-driven optimization has transformed methodologies in chemical and materials engineering, particularly in catalyst discovery, process optimization, and materials design. Ibn-Mohammed et al. proposed an AI-enabled framework for improving predictions of lifecycle environmental impacts of functional materials and devices (FM&D) [31], as shown in Figure 17. Jassim et al. developed and evaluated three supervised machine learning (ML) models—multilayer perceptron (MLP), support vector regression (SVR), and random forests (RFs)—to predict key quality parameters, including nitrification-related nitrogen compounds, in a wastewater treatment plant [97]. They further applied an iterative hyperparameter tuning algorithm to optimize the models’ performance.

With the development of Industry 4.0, artificial intelligence (AI) is gaining increasing attention for its performance in solving particularly complex problems in industrial chemistry and chemical engineering. He et al. provides an overview of the application of AI techniques, in particular machine learning, in chemical design, synthesis, and process optimization over the past years [98].

In chemical process engineering, AI algorithms optimize complex reaction parameters, separation processes, and plant-wide operations, leading to improved yields, reduced energy consumption, and enhanced safety. Nayak et al. integrated an artificial neural network (ANN) with genetic algorithms (GAs) to optimize the wastewater treatment process [99]. In their methodology, the ANN functioned as the fitness evaluator within the GA framework, enabling more accurate and adaptive optimization of process parameters, and their results demonstrated that the biomass productivity improved by 57% as compared with un-optimized conditions. Methanol produced through the power-to-liquid (PtL) process has recently gained significant attention as a promising alternative fuel. Sultan et al. presented a machine learning (ML)-based data-driven surrogate model employing artificial neural networks (ANNs) to optimize the methanol production process [100]. The results demonstrated a 33.59% increase in production rate, a 2.06% improvement in purity, and a 9.68% reduction in energy requirements compared to the baseline conditions.

These approaches can identify optimal operating conditions that might be non-intuitive to human operators, particularly for processes with complex reaction kinetics and mass transfer limitations.

The field of materials informatics has emerged as a distinct discipline, leveraging AI to accelerate the discovery and development of novel materials with tailored properties [101]. Machine learning models can predict material properties based on composition and processing parameters, guiding experimental efforts toward promising regions of the materials design space. This approach has significantly reduced the time and cost associated with materials development, particularly for applications in energy storage, catalysis, and lightweight structures. The integration of AI with high-throughput experimentation and computational modeling has created powerful workflows for materials discovery and optimization [102].

Nano-modification, which involves incorporating nanomaterials into bitumen, has emerged as a promising technique for enhancing pavement performance and durability [103]. Bhowmik et al. developed a data-driven approach for predicting and optimizing CO₂ reduction efficiency in a nano-modified bitumen system by integrating machine learning (support vector), dimensionality reduction and optimization techniques, combining genetic algorithms for global exploitation and particle swarm optimization for fine-tuning [104]. Their results demonstrated that the proposed approach achieved a CO₂ reduction of 65.738%.

5.4.3. Membrane Fouling

Membrane fouling is the process by which organic, inorganic, or biological materials accumulate on the surface or within the pores of a membrane and is widely used in water and wastewater processes. Bagheri et al. reviewed the integration of artificial intelligence (AI) techniques, such as machine learning, neural networks, and fuzzy logic, with membrane-based water and wastewater treatment processes [13]. Their work emphasized that AI-driven models can enhance process optimization, fault detection, energy efficiency, and predictive maintenance, ultimately reducing waste generation and supporting cleaner production goals. The study summarized several design frameworks and real-world case studies showing that AI can effectively model complex nonlinear relationships between operational parameters, leading to improved membrane performance, lower fouling rates, and sustainable plant operations. Al-Obaidi et al. solved the optimization of a reverse osmosis-based wastewater treatment system for the removal of chlorophenol using genetic algorithms [105]. Al-Kathiri et al. demonstrated that optimization and modeling techniques, when combined with other AI and ML algorithms, can be effectively utilized to intelligently monitor and control membrane fouling in various wastewater treatment processes, as well as during the generation of clean blue energy [106].

The selected references in Chemical and Materials Engineering, along with their key characteristics and performance metrics, are summarized in

Table 7. AI-Driven Engineering Optimization in Chemical and Materials Engineering.

Reference	Key Characteristics	Problems	Performances
Hsu et al. [35]	CNN + Deep Q-Network + reinforcement learning-based optimization	Woven composite	2.37-fold improvement in strain energy density; 267-fold speedup
Kim et al. [94]	DL + GA	Superior materials	Needs small datasets (0.5% of the initial datasets).
Nayak et al. [99]	ANN + GA	Wastewater treatment process	Improve productivity by about 57%
Sultan et al. [100]	ANN + DE	Green methanol production process	R2 is over 0.9831; 33.59% increase in production rate; 9.68% reduction in energy requirements
Bhowmik et al. [104]	SV + GA + PSO	Nano-modified Bitumen	CO2 reduction of 65.738%

.

5.5. Energy

AI-driven engineering optimization has been applied to machine learning, data analytics, and intelligent control to improve the efficiency, reliability, sustainability, and cost-effectiveness of energy systems across generation, storage, transmission, and consumption.

5.5.1. Power Energy

In power systems, AI algorithms optimize generation dispatch, grid operation, and energy trading, facilitating the integration of renewable energy sources while maintaining grid stability. Agupugo et al. summarized the application of AI and machine learning in managing microgrid systems. AI technique plays a critical role in optimizing microgrid systems, especially in forecasting equipment failures, scheduling timely repairs and then minimizing downtime and reducing maintenance costs [107]. One of the most significant impacts of AI-driven microgrid optimization on rural electrification is the cost savings [108].

Khan et al. developed a Bayesian optimization approach based on artificial neural networks to tune the voltage source converter, achieving significant improvements in dynamic response under fault conditions [109]. Biswas et al. reviewed AI-driven approaches for optimizing power consumption and evaluated the performance and outcomes of 17 distinct research methodologies, highlighting their strengths and limitations [110]. Some of these, which used AI techniques, are summarized in

Table 8. Summaries of Methodologies in Biswas et al. [110].

Reference	Key Characteristics	Problem	Outcomes
Boubaker et al. 2023 [111]	Deep Learning + Machine Learning	Photovoltaic Diagnosis and detection	Accuracy of 98.7%
Ağbulut et al. 2020 [112]	Compare ML, SVM, ANN, DL	Power output of V-trough photovoltaic system	SVM outperforms with R2 of 0.9921
Kannari et al. 2023 [113]	Reinforcement learning (RL)	Heating cost	Cost reduction by 23%
Yang et al. 2020 [114]	Machine learning	Building heat consumption	36–38% of energy saving
Ahmed et al. 2020 [115]	Machine learning	Energy management of smart grid
Malta et al. 2023 [116]	Reinforcement learning	Management of 5G base stations	75% energy saving with 20 ms

.

Ashraf et al. proposed a framework, as shown in Figure 18, incorporating artificial intelligence (AI) for improving the efficiency of a high-pressure steam turbine [117]. Two AI models, artificial neural network (ANN) and support vector machine (SVM), are used to train data and to generate an optimization model and a Monte-Carlo -based method is used to optimize the efficiency of the turbine. Their results showed that the turbine efficiency can be improved by 3.4% compared to the average values of full-load power generation models.

5.5.2. Renewable Energy

With the rapid increase in global energy consumption, renewable energy has received special attention globally due to climate change and ecological issues. The “2025 Global Offshore Wind Report” [118] shows that in 2024, 56.3 GW of offshore wind capacity was added, with 23.2 GW in Europe and 17.4 GW in China, bringing the total global offshore wind power installation capacity to 83 GW.

Figure 19 shows a schematic of various forms of renewable energy in which AI is used for design, optimization, control and so on [108,119]. Most AI techniques are used with neural networks to forecast wind speed and wind strength in the area of renewable energy [120]. ANN is one of the most widely used models in the last decade, which consists of many layers, one input layer and one output layer. Mohandes et al. compared the results from the ANN model and the Autoregressive (AR) Model for wind speed prediction. The result indicated that the ANN was superior to the AR model for daily and monthly mean speed values [121]. Mabel et al. applied feedback for the prediction of wind strength over a span of time to feed a backpropagation neural network (BPNN) [122]. Their results showed that the root mean square errors in the BPNN for training and evaluation are 0.0070 and 0.0065, respectively. Flores et al. developed an optimization and control algorithm for wind speed and active power prediction based on an ANN model using the back-propagation method and presented a Genetic algorithm for the active power generation of a wind farm [123].

The Maximum Power Point Tracking (MPPT) problem in wind turbine control is used to optimize turbine performance and enhance the system’s power generation efficiency by using evolutionary algorithms [119]. Evolutionary algorithms are widely recognized as powerful optimization tools. For instance, the genetic algorithm (GA) has been used by Guediri et al. to adjust FLC (fuzzy logic control) system parameters for optimizing wind turbine MPPT strategy [124], and the Fuzzy Logic control strategies for the offshore wind turbine [125]. A multi-objective particle swarm optimization (MPSO) had been applied to optimize the control parameters of yaw control systems in horizontal-axis wind turbines, aiming to improve energy capture efficiency [126]. The yin–yang grey wolf optimizer (YYGWO) algorithm developed by Mirjalili et al. [127], through nonlinear model predictive control, has been employed for maximizing wind energy extraction in large wind turbines [126]. Neural Networks are primarily used for model prediction and system behavior simulation in MPPT applications. For instance, an unsupervised neural network has been used to develop a direct speed control (DSC) strategy model of wind turbines to maximize the generated output power [128]. Muñoz-Palomeque et al. [128] demonstrated that the proposed neural-based strategy model can provide around 1 7.87% more power. Sun et al. [129] developed power prediction models for wind turbines using artificial neural network models, and optimized yaw angles across wind farms to reduce wake effects and enhance overall efficiency. The power ratio of wind turbines can reach 0.96 in all directions.

Solar energy production is highly dependent on meteorological conditions such as sunlight intensity, cloud cover, and temperature. Mellit and Kalogirou showed that ANNs could outperform traditional statistical methods in predicting solar irradiance, which directly impacts photovoltaic (PV) system output [130].

The integration of renewable energy sources (RESs) such as wind and solar into existing power grids introduces complex challenges primarily because of their intermittent and unpredictable nature. Hossain et al. demonstrated that AI has a high potential and applicability for decentralized energy grids [131]. Belrzaeg et al. applied artificial intelligence to optimize renewable energy grids to enhance their efficiency and sustainability [132]. The AI control unit is used to optimize charging and discharging cycles of energy storage systems (see Figure 20).

Hydrogen is widely recognized as a key energy carrier for a sustainable future, thanks to its ability to store and deliver energy in a clean and efficient manner. Bhuiyan et al. presented a comprehensive review of artificial intelligence (AI)-driven optimization in renewable hydrogen production [133]. Bassey et al. applied machine learning to train more accurate and efficient models of identifying patterns and relationships of data from various stages of the lifecycle of renewable energy systems, including raw material extraction, manufacturing, installation, operation and maintenance [134]. AI-driven diagnostics can detect early signs of anomalies and diagnose faults in fuel cell components, providing recommendations for remedial actions in real-time and extending the lifespan of hydrogen-powered systems [135].

5.5.3. Food Dryer System

As the primary method of fresh food preservation, the drying technique has proven to have many advantages, such as permitting early harvest, reducing shipping weights and costs and minimizing packaging [136,137]. Sun et al. reviewed recent developments of artificial intelligence in drying of fresh food [138]. Nadian et al. developed an intelligent fuzzy-machine control system, in which two separated neural networks were used to predict food material property and moisture ratio and a genetic algorithm was applied to find the optimal control parameters [33]. Their results had demonstrated that the new system could significantly reduce the drying time and had a good balance between energy consumption and food quality.

5.5.4. Unmanned Aerial Vehicles

Haider et al. combined clustering techniques to reduce the energy consumption of unmanned aerial vehicles (UAVs) through meticulously planned routing and path determination [139]. Their results demonstrated an impressive 15% increase in energy conservation and a 20% reduction in data transmission time.

The selected references in Energy, along with their key characteristics and performance metrics, are summarized in

Table 9. AI-Driven Engineering Optimization in Energy.

Reference	Key Characteristics	Problems	Performances
Khan et al. [109]	ANN + Bayesian optimization	Power Converters	Significant improvements in dynamic response
Ashraf et al. [117]	ANN + SVM + Monte-Carlo -based method	High-pressure steam turbine	Efficiency improved by 3.4%
Mohandes et al. [121]	ANN + Autoregressive (AR) Model	Wind speed prediction	ANN was superior to AR model
Mabel et al. [121]	backpropagation neural network	Prediction of wind strength	MSE is 0.007
Flores et al. [123]	ANN + GA	Wind speed and active power prediction
Guediri et al. [124]	GA	Wind Power System	Greater efficiency, impressive results,
Song et al. [126]	Yin-Yang grey wolf optimization	Large-scale wind turbines	Capture 0.03–0.04% energy
Muñoz-Palomeque et al. [128]	NN	Wind turbine	7.87% more power
Sun et al. [129]	ANN	Wind turbines	The power ration can reach 0.96
Nadian et al. [34]	GA	Tybrid hot air–infrared dryer	Energy consumption (0.158 kW h)
Haider et al. [139]	clustering technique cluster-based dynamic algorithm	Unmanned aerial vehicles	15% increase in energy conservation; 20% reduction in data transmission time

.

5.6. Managements

AI-driven engineering optimization has increasingly influenced management science by enabling data-driven, adaptive, and intelligent decision-making in complex organizational systems.

5.6.1. Supply Chains

As global supply chains become increasingly complex, the importance of effective supply chain risk management has grown substantially. The entire economy of a country largely depends on an efficient and well-managed supply chain process, and genetic algorithms are a common method used to solve supply chain management problems [140,141,142]. Grover explored the application of artificial intelligence (AI) in managing supply chain risks, examining recent advancements, existing challenges, and potential directions for future research [143].

Mass Customization manufacturing, as one supply chain problem, involves high-volume production with a wide variety of materials, has received growing consideration by researchers and practitioners since 1980 [144]. Alfayoumi et al. built a multi-objective mass customization optimization problem to minimize the time and cost of mass customized orders by using the NSGAII algorithm [145]. They claimed that the time had been improved by 20.4% and the cost had been reduced by 29.8% compared to traditional expert optimization.

The downstream petroleum sector faces significant challenges in transport and distribution logistics, particularly for small and medium enterprises (SMEs) that rely on efficient supply chains to remain competitive. Abudu & Sai presented the crude oil value chain/system shown in Figure 21 [146]. Arinze et al. proposed and implemented an AI-driven Transport and Distribution Optimization Model (TDOM) for the downstream petroleum sector [147]. Machine learning algorithms are incorporated into the TDOM model to optimize dynamic routing and scheduling. Supervising techniques are used to predict delivery times, identify bottlenecks and recommend optimal routines. Unsupervised learning allows the model to detect anomalies in delivering patterns and identify hidden inefficiencies that may not be immediately identified through traditional methods [148]. Arinze et al. demonstrated that by using the TDOM, SMEs experienced an average 15% reduction in fuel costs, a 20% improvement in on-time delivery rates, and a 10% decrease in overall transportation expenses [147]. Arinze et al. also identified some difficulties and challenges, such as (1) the integration of diverse data sources, particularly from SMEs that used different systems for inventory management, fleet tracking and customer orders, and (2) the initial resistance from some SMEs to adopting AI-driven technology.

The fashion industry’s significant contribution to global pollution underscores the urgent need for a systemic shift toward sustainability, requiring innovation in materials, transparency in supply chains, and changes in consumer behavior [149]. Donthi et al. investigated how AI-driven optimization can reduce the industry’s carbon footprint by improving supply chain efficiency, 25% reduction in emissions, and resource management with linear programming, genetic algorithms and reinforcement learning [150].

The agricultural sector, facing the dual challenges of meeting the growing food demands of a burgeoning global population and adapting to the complexities of climate change, requires innovative solutions such as precision agriculture, data-driven decision-making, and AI-assisted resource management to ensure long-term food security [151]. Elufioye et al. made a comprehensive review on the application of AI in agricultural supply chains via real-time data analysis, predictive maintenance and resource optimization [151]. Espolov et al. discuss a methodology for the efficient utilization of economic resources through supply chain optimization within the Asian agricultural market [152].

E-commerce supply chains have faced significant challenges, including rising consumer demand, heightened competitive pressure, and the critical need to ensure seamless operational efficiency. Kaul and Khurana investigated advanced AI-driven techniques, such as neural networks, deep reinforcement learning, and optimization algorithms, to enhance real-time responsiveness, reduce operational costs, and strengthen overall supply chain resilience [153].

5.6.2. Maintenance

Hamdan et al. reviewed the state of artificial intelligence applications in the renewable energy domain, especially in modelling predictive maintenance and energy optimization across diverse sources such as solar, wind, and hydro [154]. Bello et al. employed a combination of machine learning algorithms, including deep neural networks and reinforcement learning, to develop predictive maintenance models, trained on large-scale datasets from operational wind farms, solar installations and hydroelectric plants [155]. Their results showed that the AI-driven techniques can predict equipment failures with 92% accuracy and reduce unplanned downtime by 35% compared to traditional methods.

5.6.3. Project & Process

Project management is on the cusp of a major transformation, driven by the integration of Artificial Intelligence (AI). Bhanderi explored the transformative role of Artificial Intelligence (AI) in modern project management, emphasizing how AI technologies are reshaping decision-making, workflow optimization, and resource management [156]. By examining current applications, benefits, and challenges, Bhanderi also demonstrated how AI-driven approaches are enabling organizations to enhance efficiency, reduce risks, and achieve greater project success [156]. Li integrated reinforcement learning and predictive analysis techniques to optimize resource allocation and task scheduling, achieving an average 31.2% increase in resource utilization and a 24.8% reduction in operational costs [157].

The integration of artificial intelligence (AI) into manufacturing has transformed quality control and process optimization. Okuyelu and Adaji reviewed recent advances in AI technologies, focusing on their applications in manufacturing [158]. They also developed an AI-based real-time monitoring system that continuously tracks production parameters and improves manufacturing performance through real-time analytics, adaptive control, predictive maintenance, and intelligent decision-making.

5.6.4. Inventory & Logistics

Many airlines continue to face challenges with imbalances in spare parts inventories, leading to either costly excesses or critical shortages that disrupt maintenance schedules and operational readiness. MoghadasNian examined the application of AI-driven predictive analytics to optimize inventory management in airline logistics, aiming to reduce spare part shortages, enhance operational efficiency, and indirectly contribute to sustainability by minimizing waste and energy consumption [159]. The study reported cost reductions ranging from 25% to 40% and inventory level decreases of 20% to 54%.

The logistics and freight industry plays a crucial role in the global economy by facilitating the transportation of goods from producers to consumers. Royappa et al. discussed the applications of AI in freight and logistics management to enhance efficiency and sustainability [160]. Their study demonstrated that AI-driven optimization can reduce operational costs by 17% and improve customer satisfaction levels by 22%. Thuraka developed an adaptive routing framework modeled using a deep Q-learning algorithm and solved through a Markov decision process [161]. He also highlighted several implementation challenges, including data inconsistency, system interoperability issues, and the need for supportive policies.

The growing complexity of urban logistics—driven by rapid urbanization and the surge in e-commerce—demands intelligent and sustainable routing strategies. Mahat et al. demonstrated that ant colony optimization (ACO) can address various optimization challenges in supply chain and logistics problems [162]. ACO can determine the most efficient routes to improve customer satisfaction through on-time deliveries and reduced costs. Their results showed that ACO outperforms genetic algorithms, simulated annealing, and linear programming approaches.

5.6.5. Manufacturing Process

Osho et al. proposed a conceptual framework for AI-driven predictive optimization in industrial engineering, focusing on the use of machine learning (ML) algorithms to enhance decision-making in smart manufacturing [163]. The framework integrates real-time data acquisition, advanced data preprocessing, predictive analytics, and optimization layers to enable proactive and adaptive decision-making across manufacturing processes.

The selected references in management, along with their key characteristics and performance metrics, are summarized in

Table 10. AI-Driven Engineering Optimization in Management.

Reference	Key Characteristics	Problems	Performances
Alfayoumi et al. [145]	NSGAII	Mass customized orders	The time improved by 20.4% and the cost reduced by 29.8%
Arinze et al. [147]	ML + route optimization	Downstream petroleum sector–Supply Chain	20–30% reduction in transport costs
Donthi et al. [150]	RL + GA	Fashion industry’s supply chain	25% reduction in emissions
Kaul and Khurana [153]	ML + Optimization	E-commerce Supply Chain	Cost efficiency
Bello et al. [155]	ML (DL + RL)	Wind farm maintenance	Failures with 92% accuracy
Li [157]	RL + scheduling optimization	Resource allocation	31.2% increase in resource utilization and 24.8% reduction in operational costs
MoghadasNian [159]	AI + Optimization	Airline logistics	Cost reductions from 25% to 40%; inventory decreases of 20% to 54%.
Royappa et al. [160]	ML + route optimization	Freight and logistics management	Cost reduction by 17%; customer satisfaction increases by 22%.

.

6. Implementation Challenges and Limitations

6.1. Data Quality and Availability

The performance of AI-driven optimization approaches is fundamentally constrained by the quality, quantity, and relevance of available training data. Engineering applications often face the challenge of data scarcity, particularly for rare failure events or extreme operating conditions that are critical for robust optimization. Furthermore, engineering data frequently exhibits non-uniform quality, with inconsistencies arising from different measurement techniques, sensor calibrations, or environmental conditions. These data challenges can lead to biased models, inaccurate predictions, and suboptimal solutions when not properly addressed.

The curse of dimensionality presents another significant challenge in data-driven engineering optimization. As the number of design variables increases, the volume of data required for accurate model training grows exponentially, making comprehensive data collection prohibitively expensive or time-consuming for complex engineering systems. This challenge is particularly acute in fields relying on physical experiments or high-fidelity simulations, where data generation is resource-intensive. Techniques such as active learning, transfer learning, and multi-fidelity modeling have shown promise in addressing these limitations by strategically selecting the most informative data points and leveraging correlations between different levels of model fidelity.

6.2. Model Integration

The integration of AI models into established engineering workflows presents significant technical and cultural challenges. Many engineering organizations rely on validated simulation tools and establish design practices that may be difficult to reconcile with data-driven approaches. The black-box nature of many complex AI models, particularly deep neural networks, creates barriers to adoption in safety-critical applications where model interpretability and traceability are essential requirements. Engineers rightly question optimization results that lack physical justification or transparent reasoning.

6.3. Computational Demands and Scalability

While AI-driven optimization often aims to reduce computational requirements compared to traditional approaches, the training of complex AI models itself demands substantial computational resources. Deep neural networks with millions of parameters require significant memory, processing power, and time for training, particularly when dealing with high-dimensional engineering problems. The hyperparameter tuning process further exacerbates these computational demands, often requiring multiple training runs to identify optimal model configurations.

The scalability of AI-driven optimization methods presents another significant challenge, particularly when moving from academic test cases to industrial-scale applications. Engineering systems in practice often involve coupled physics, multiple scales, and numerous design variables that strain current AI methodologies. Techniques such as domain decomposition, multiscale modeling, and distributed computing have shown promise in addressing these scalability challenges. Additionally, the development of specialized hardware for AI workloads, including graphics processing units (GPUs) and tensor processing units (TPUs), continues to push the boundaries of feasible problem sizes for AI-driven optimization.

6.4. Explainability

Explainability in an AI system allows highlighting the relevant information in order to make the best-informed decisions. Explainability allows verifiability of decisions (as long as the resulting explanations are sufficiently informative [164], which is particularly important as contemporary models often underperform in generalization and out-of-distribution samples, of which there are plenty of instances in the engineering and manufacturing settings.

Recent research has made substantial progress in enhancing the interpretability and explainability of AI models for engineering applications. Techniques such as attention mechanisms, feature importance analysis, and surrogate interpretable models provide insights into model behavior and decision processes. Furthermore, the integration of physical principles directly into AI architecture through physics-informed neural networks or hybrid modeling approaches enhances model plausibility and alignment with engineering knowledge. These developments help bridge the gap between data-driven predictions and engineering intuition, facilitating broader adoption of AI-driven optimization in practice.

Explainability has been applied within engineering, although its adoption remains limited in practice [165]. In processes such as injection molding, defects are rare, leading to highly imbalanced datasets that make it difficult to fine-tune process parameters. Shapley Additive Explanations (SHAP) [166] has been used in this context to identify the process variables most responsible for defect occurrence [167]. SHAP assigns each input feature a contribution score to a given prediction, based on concepts from cooperative game theory. In automated visual inspection of industrial products [168], a Double-VGG16 CNN to classify defect types and then apply Grad-CAM [169] to localize the defects, checking where the network ‘looked’ when labelling a scratch or pit.

Some approaches add model interpretability or model visualization, which are useful in designing model architectures or tuning hyperparameters, but not for providing high-level explanations to end-users. The intersection of domain knowledge with explainability is an active research area, for instance, validating ML decisions against causal knowledge graphs [170].

7. Outlook

7.1. Emerging Trends

7.1.1. Physics-Informed AI and Scientific Machine Learning

The integration of physical principles with data-driven AI approaches represents one of the most promising directions for engineering optimization. Physics-Informed Neural Networks (PINNs) and related architectures embed fundamental conservation laws, constitutive relationships, and boundary conditions directly into the learning process, ensuring physical consistency while reducing data requirements. These approaches are particularly valuable for engineering applications where data may be sparse, noisy, or expensive to acquire, but where physical principles are well-established.

Beyond PINNs, the emerging field of scientific machine learning is developing novel architectures specifically designed for engineering and scientific applications. Geometry-aware neural networks can handle complex engineering geometries without resorting to simplistic pixelation approaches that lose information near boundaries. Operator learning frameworks aim to learn mappings between function spaces, enabling generalization across different geometries, boundary conditions, and system parameters. These advances promise to enhance the robustness, accuracy, and generalizability of AI-driven engineering optimization, particularly for complex multi-physics systems.

7.1.2. Human-AI Collaboration and Interactive Optimization

Future engineering optimization frameworks will increasingly emphasize collaborative intelligence that leverages the complementary strengths of human engineers and AI systems. Rather than replacing human expertise, AI will serve as an augmentation tool that enhances human capabilities in conceptual design, decision-making, and creative problem-solving. Interactive optimization approaches will allow engineers to guide the search process based on experience and intuition while leveraging AI to handle computational complexity and identify non-obvious relationships.

The development of explainable AI (XAI) techniques specifically tailored for engineering applications will be crucial for facilitating effective human–AI collaboration. Visualization tools that illustrate the reasoning behind AI recommendations, uncertainty quantification methods that communicate confidence levels, and interface designs that enable natural interaction with AI systems will help bridge the gap between data-driven optimization and engineering judgment. These approaches will be particularly valuable in scenarios requiring trade-off decisions between competing objectives, where human values and contextual knowledge play essential roles in the optimization process.

7.1.3. Edge AI and Real-Time Optimization

The convergence of AI with edge computing enables real-time optimization capabilities for engineering systems with strict latency requirements or limited connectivity. TinyML approaches allow compact AI models to be deployed directly on embedded systems, sensors, and controllers, enabling localized decision-making without continuous cloud connectivity. These capabilities are particularly valuable for applications such as autonomous systems, industrial robotics, and adaptive infrastructure, where rapid response to changing conditions is essential.

Real-time optimization through edge AI also presents significant opportunities for autonomous systems and adaptive infrastructure. In manufacturing, AI-driven real-time optimization can adjust process parameters to maintain quality despite material variations or environmental fluctuations. In civil infrastructure, embedded AI systems can optimize structural control systems in response to changing loads or environmental conditions. These applications demonstrate the potential for AI-driven optimization to move beyond the design phase and into the operational lifecycle of engineering systems, enabling continuous performance improvement and adaptation.

7.2. Environmental Sustainability Considerations

Ensuring that digital technologies meet current operational needs without jeopardizing the welfare and opportunities of future generations is a central principle of environmental sustainability [171]. When viewed through this lens, AI systems are not impact-free. Their development, deployment, and eventual retirement draw on substantial natural and material resources.

7.2.1. Resource Use Across the AI Lifecycle

AI technologies rely on a broad range of inputs, including electricity, cooling water, and specialized minerals used in semiconductors and data-center equipment. These resources are consumed directly during the manufacturing, operation, and disposal of hardware, and indirectly through processes such as electricity production and infrastructure maintenance. The cumulative effects of this resource use—air and water pollution, thermal emissions, and the accumulation of electronic waste—can pose risks to ecosystems and human wellbeing. Because these impacts are not evenly distributed globally, they may also intensify existing inequalities or contribute to new forms of social disadvantage [172,173].

7.2.2. Emerging Trends in Environmental Impact

Although the long-term environmental footprint of AI is still evolving, current patterns in the field point to several clear trajectories:

Rising computational demands for training: Contemporary machine learning models require significantly greater volumes of data and computing power compared with earlier generations. The computational intensity of leading models, particularly for complex multiscale systems, has risen dramatically over the past decade [174,175].

Increasing cost of inference: The expanding scale of models—reflected in higher parameter counts and larger input sizes means that day-to-day use is becoming more resource-intensive [176,177].

Growth in data-storage requirements: Global storage capacity in data centers and user devices is expected to more than double between 2023 and 2027, with AI-driven data generation being a significant contributing factor [178,179].

Escalating need for advanced hardware: Demand for high-performance chips and related semiconductor technologies is forecast to continue accelerating. Sales of processors designed specifically for generative AI could exceed $400 billion by 2027 [180,181].

Together, these trends indicate that the energy, water, and material demand associated with AI lifecycles are increasing and are likely to rise even further.

Electricity consumption is a particular area of concern. Some projections warn that, in certain regions, AI-related electricity usage could exceed renewable energy supply within the next decade. Additional modelling suggests that rapid growth in data-center energy use—largely driven by AI workloads—may push total energy demand in the United States beyond its projected generating capacity as early as 2028 [182].

7.2.3. The Need for Lifecycle-Sustainable AI

AI has the capacity to deliver substantial environmental, economic, and social value, but these benefits can only be realized if the systems themselves are developed and operated sustainably. Without careful management, the resource demands of AI could reinforce environmental degradation and hinder progress toward renewable energy integration.

Addressing these risks requires coordinated action across the full AI value chain—from model design and infrastructure planning to deployment and end-of-life management [171]. Early and proactive efforts to measure, understand, and reduce resource use are essential to ensuring that AI contributes positively to long-term sustainability goals rather than undermining them.

7.2.4. Environmentally Friendly AI Methods

Environmentally friendly AI engineering optimization methods focus on lowering energy use, computational costs, and environmental impact while ensuring reliable performance. One approach is energy-aware neural architecture search, where AI models are designed and chosen not only for accuracy but also for low energy consumption and quick training times. Lightweight, high-precision models, such as reduced-order models, surrogate models, and physics-informed machine learning, can replace costly simulations and reduce computational demands significantly. Techniques like model pruning, quantization, and knowledge distillation shrink the size and energy requirements of models without losing performance. Transfer learning and incremental learning decrease energy consumption further by reusing existing models instead of training new ones. Also, life cycle assessment can be included in AI evaluation to track environmental impact across the data collection, training, deployment, and hardware usage stages. Ethical and responsible AI practices are also key and should include bias detection using fairness metrics, transparent and explainable models to support accountability, and human-in-the-loop decision-making to guarantee AI-driven engineering decisions are fair, trustworthy, and aligned with societal and environmental values.

8. Conclusions

This comprehensive review has examined the transformative impact of AI-driven optimization across engineering disciplines, highlighting fundamental methodologies, domain-specific applications, implementation challenges, and emerging research directions. The integration of artificial intelligence into engineering optimization represents a paradigm shift with far-reaching implications for how engineering systems are designed, operated, and maintained. AI technologies have demonstrated remarkable capabilities in navigating complex design spaces, reconciling competing objectives, and identifying non-intuitive solutions that might elude human designers or traditional optimization approaches.

Despite significant progress, substantial challenges remain in data quality, model interpretability, computational demands, and real-world implementation. Addressing these challenges requires continued research at the intersection of AI and engineering, the development of standardized evaluation frameworks, and the creation of educational programs that prepare the next generation of engineers with both domain expertise and AI proficiency. The most promising future directions involve tighter integration of physical principles with data-driven approaches, enhanced human–AI collaboration, and real-time optimization capabilities enabled by edge computing.

As AI technologies continue to evolve and mature, their role in engineering optimization will undoubtedly expand, enabling more sustainable, efficient, and innovative engineering solutions to society’s most pressing challenges. By embracing the opportunities while thoughtfully addressing the limitations and risks, the engineering community can harness the full potential of AI-driven optimization to advance human capability and improve quality of life worldwide.