Atomistic Modeling of Microstructural Defect Evolution in Alloys Under Irradiation: A Comprehensive Review

Abstract

Developing structural materials capable of maintaining integrity under extreme irradiation conditions is a cornerstone challenge for advancing sustainable nuclear energy technologies. The complexity and severity of radiation-induced microstructural changes—spanning multiple length and timescales—pose significant hurdles for purely experimental approaches. This review critically evaluates recent advancements in atomistic modeling, emphasizing its transformative potential to decipher fundamental mechanisms driving microstructural evolution in irradiated alloys. Atomistic simulations, such as molecular dynamics (MD), have successfully unveiled initial defect formation processes at picosecond scales. However, the inherent temporal limitations of conventional MD necessitate advanced methodologies capable of exploring slower, thermally activated defect kinetics. We specifically traced the development of powerful potential energy landscape (PEL) exploration algorithms, which enable the simulation of high-barrier, rare events of defect evolution processes that govern long-term material degradation. The review systematically examines point defect behaviors in various crystal structures—BCC, FCC, and HCP metals—and elucidates their characteristic defect dynamics, respectively. Additionally, it highlights the pronounced effects of chemical complexity in concentrated solid-solution alloys and high-entropy alloys, notably their sluggish diffusion and enhanced defect recombination, underpinning their superior radiation tolerance. Further, the interaction of extended defects with mechanical stresses and their mechanistic implications for material properties are discussed, highlighting the critical interplay between thermal activation and strain rate in defect evolution. Special attention is dedicated to the diverse mechanisms of dislocation–obstacle interactions, as well as the behaviors of metastable grain boundaries under far-from-equilibrium environments. The integration of data-driven methods and machine learning with atomistic modeling is also explored, showcasing their roles in developing quantum-accurate potentials, automating defect analysis, and enabling efficient surrogate models for predictive design. This comprehensive review also outlines future research directions and fundamental questions, paving the way toward autonomous materials’ discovery in extreme environments.

1. Introduction

1.1. The Grand Challenge of Materials in Extreme Environments

The rapid expansion of global energy demands, combined with increasing concerns over carbon emissions and climate change, has revitalized interest in nuclear power as a critical component of sustainable energy strategies. Advanced nuclear reactor concepts, including Generation IV reactors and fusion reactors, promise enhanced safety, efficiency, and sustainability. However, these advanced reactors typically operate under conditions of extreme irradiation, high temperatures, and aggressive chemical environments, placing unprecedented demands on structural materials [1,2,3,4,5]. One of the central challenges in enabling the deployment of next-generation reactors is developing radiation-tolerant materials that maintain structural integrity and functionality over extended periods under intense irradiation. Irradiation creates a highly dynamic and complex microstructural landscape characterized by defects such as vacancies, self-interstitial atoms (SIAs), defect clusters, and more extended and mechanically consequential defects such as dislocation loops. The formation, evolution, and interactions of these defects drive significant changes in the material’s microstructure, resulting in phenomena such as swelling, hardening, embrittlement, creep, etc. [6,7,8,9,10,11,12,13,14].

Historically, understanding the irradiation-induced degradation mechanisms has been largely empirical, relying extensively on experimental observations obtained from irradiation facilities and reactor operation data. These experimental studies, while crucial, have inherent limitations. Directly observing atomic-scale defect phenomena experimentally is extremely challenging, if not impossible, due to their transient nature and the limitations of resolution and timescales achievable by characterization techniques such as electron microscopy. Furthermore, irradiation-induced defects evolve across multiple length scales—from sub-nanometer point defects to micrometer-sized voids and dislocation loops—and across vastly different timescales, from picoseconds during displacement cascades to years or decades of reactor operation. This multiscale complexity makes it exceptionally challenging to predict material behavior reliably and quantitatively solely through experimental means. The nuclear materials community has traditionally addressed these challenges through iterative cycles of experimentation, empirical modeling, and alloy optimization. However, this trial-and-error approach is both costly and time-consuming, hindering the timely development of novel reactor technologies. Therefore, there is a critical need for comprehensive computational modeling approaches that can complement experimental studies, provide atomic-level insights, and predict long-term material behavior under extreme irradiation conditions.

Advanced computational methodologies, particularly atomistic modeling, offer a promising solution to probe the fundamental mechanisms governing defect formation, migration, clustering, and annihilation processes at atomic resolutions. These computational methods have the potential to dramatically accelerate material design cycles by providing quantitative predictions of irradiation-induced microstructural evolution, guiding the development of new alloys tailored for superior irradiation resistance. This review focuses specifically on the role of atomistic modeling in addressing the grand challenges posed by irradiation effects in alloys. We critically assess the methodological advancements and demonstrate how these powerful computational tools can contribute to resolving the complex interplay of defects under irradiation, ultimately enabling the informed design of materials resilient in extreme reactor environments.

1.2. The Promise of Atomistic Modeling

Atomistic modeling encompasses a suite of simulation techniques that explicitly consider the collective behavior of constituent atoms under prescribed physical conditions. This modeling paradigm has revolutionized materials science by enabling researchers to dissect the atomic-level mechanisms that underpin macroscopic properties. In the context of irradiation, where primary damage and subsequent defect evolution occur at the atomic scale, such models offer unparalleled insight into processes that are otherwise inaccessible.

The primary advantage of atomistic simulations lies in their ability to directly simulate fundamental physical events, such as displacement cascades, defect migration, and recombination. Unlike continuum approaches, atomistic models do not rely on empirical constitutive relations; instead, they build upon interatomic potentials—which are usually calibrated with first-principles calculations—to produce physically grounded predictions. This capability is particularly important in radiation damage studies, where the sequence of atomic collisions during a cascade, the formation of stable and metastable defect structures, and their subsequent dynamics are too complex to be described in an analytic manner. It is worth mentioning that, arguably, one of the earliest atomistic simulations was employed to probe the radiation damage in metals [15].

One of the most widely used atomistic techniques in this area is classical molecular dynamics (MD), which integrates Newton’s equations of motion to simulate atom trajectories. MD has been instrumental in understanding the morphology of displacement cascades in various metals, the formation of defect clusters, and the emergence of dislocation loops. However, conventional MD is inherently limited by the accessible simulation timescales (typically up to nanoseconds) and system sizes (usually 10³~10⁶ atoms). The length scale challenge can be alleviated by the parallelization computing techniques. Nowadays, the state-of-the-art MD simulations can handle more than a billion atoms [16,17,18]. However, due to the arrow of time, the parallelization cannot be applied to the time domain. Hence, the conventional MD simulations still run short for capturing the full temporal evolution of defects under reactor conditions.

To overcome these limitations, a range of advanced atomistic approaches have been developed, based on the school of thought on energy landscape probing. In particular, energy landscape exploration enables the identification of stable configurations, transition states, and reaction pathways associated with defect migration and interaction. Some specific algorithms, including accelerated dynamics methods, such as temperature-accelerated dynamics (TAD) [19,20], hyperdynamics [21,22,23], the dimer method [24,25,26], the activation–relaxation technique (ART) [27,28,29], and autonomous basin climbing (ABC) [30,31] and its extended variant (ABC-E) [32], allow for the exploration of rare-event dynamics that govern long-term defect migration and transformation. These techniques have made it feasible to simulate microstructural evolution over timescales approaching seconds or longer, offering insights into processes like void swelling, phase transformations, dislocation–obstacle interactions, and grain coarsening that are critical to material degradation under irradiation. Beyond purely structural modeling, recent efforts have incorporated chemical effects, particularly in multi-component alloys and high-entropy systems, by integrating atomistic simulations with advanced interatomic potentials that reflect complex local chemistry [33,34,35,36,37,38]. These efforts are crucial for understanding how compositional heterogeneity influences defect behavior, especially in emerging alloy systems designed for extreme environments.

Atomistic modeling also plays a critical role in supporting multiscale modeling frameworks. The parameters and insights derived from atomistic simulations can be fed into mesoscale and continuum-level models to enable predictive capabilities across lengths and timescales [7,39,40,41,42,43,44,45]. For example, advanced atomistic methods (detailed below in Section 2.2) can precisely probe defect evolution/interaction mechanisms and calculate their corresponding activation energy barriers. Such data serves as direct input for kinetic Monte Carlo (KMC) simulations, which can then model the collective evolution of millions of defects over seconds or minutes that can be compared with experimental characterizations. Similarly, the stress required to move a dislocation past an obstacle, as determined by atomistic simulations (detailed below in Section 4.2), can inform the constitutive rules used in dislocation dynamics (DD) models to predict plastic flow and hardening at the scale of entire grains. Insights into interfacial energy and mobility can also parameterize phase field models that describe the evolution of microstructural features like grain growth or phase transformation. By providing this critical, bottom–up information that atomistic modeling bridges the gap between fundamental physics and the engineering-scale phenomena that govern material lifetime. This hierarchical integration is central to the development of physics-based models for radiation damage that are both accurate and computationally tractable.

In summary, atomistic modeling serves as a foundational pillar in the effort to understand and predict material behavior under irradiation. By capturing the fundamental processes at the atomic level and bridging them to macroscopic phenomena, these methods provide a powerful toolkit for designing the next generation of radiation-tolerant materials. The following sections of this review will delve into the specific methodologies, applications, and insights that atomistic simulations have contributed to this critical field of materials science.

2. Methodological Advances: From Conventional MD to Long-Term Atomistic Modeling

2.1. Classical Molecular Dynamics (MD) in Irradiation Studies

Classical MD serves as a foundational computational tool in the study of radiation effects, providing unparalleled insights into the initial creation of damage. By integrating Newton’s equations of motion for thousands to billions of atoms, MD simulations can explicitly track the atomic trajectories during a high-energy displacement cascade. This capability has been instrumental in revealing the intricate, sub-picosecond processes of defect production in a wide range of metals and alloys. Simulations have consistently shown that cascades produce a large number of vacancies and interstitials, which subsequently recombine or aggregate into more complex defect structures, such as clusters and dislocation loops, within the first few picoseconds after the primary knock-on event (e.g., Figure 1). These highly detailed, atom-level views of primary damage formation are inaccessible to direct experimental observation and thus represent a unique contribution of MD to the field.

Despite its success in modeling primary damage, conventional MD faces a severe, inherent limitation: the accessible timescale. The use of femtosecond-scale integral time steps, necessary to capture atomic vibrations, restricts typical MD simulations to total times of nanoseconds at most. However, the microstructural evolution that governs long-term material degradation under irradiation is driven by thermally activated, rare-event kinetic processes—such as defect diffusion and interaction—that unfold over timescales of seconds, days, or even decades of reactor operation. This discrepancy of more than ten orders of magnitude constitutes a formidable challenge for predictive modeling. Because MD simulations are ill suited to capture these infrequent events, they cannot directly model the slow, diffusive evolution of the defect population that leads to macroscopic phenomena (e.g., void swelling, creep, etc.). This temporal constraint forces such studies to be conducted at either athermal (0 K) conditions or unphysically high strain rates, which can fail to capture the critical interplay between thermal activation and mechanical or irradiation-driven forces, potentially leading to misleading conclusions about material behavior under realistic service conditions. Therefore, while MD is indispensable for understanding damage creation, more advanced computational methods are required to bridge the timescale gap and predict the long-term performance of irradiated materials.

It is worth noting that, while MD simulations offer profound insights, the accuracy of their predictions is fundamentally governed by the quality of the underlying interatomic potential. Traditional potentials, such as those based on the Embedded Atom Method (EAM), are computationally efficient but are often developed by fitting to a limited set of experimental or quantum mechanical data, such as lattice parameters, elastic constants, and formation energies. As a result, a significant challenge arises from a lack of transferability. A potential that accurately describes bulk mechanical properties might fail to correctly predict the energetics of point defects or the complex, distorted atomic environments within a displacement cascade core. This is particularly problematic in chemically complex multi-component alloys, where a single potential must capture the interactions between many different elements, a task for which traditional potentials often struggle. The choice of potential can therefore strongly influence simulation outcomes, from the number of defects produced in a cascade to their subsequent mobility. This limitation provides a direct and compelling motivation for the development of Machine Learning Interatomic Potentials (MLIPs), which learn the potential energy surface from vast quantum–mechanical datasets and offer near-DFT accuracy at a fraction of the computational cost, as discussed below in Section 5.1.

2.2. Beyond Conventional MD: Probing Long-Timescale Kinetics

To overcome the inherent timescale limitations of conventional MD, a number of advanced computational techniques have been developed, grounded in the concept of exploring the system’s potential energy landscape (PEL). Rather than tracking every atomic vibration, these methods focus on identifying the kinetics of rare, thermally activated events that govern long-term microstructural evolution. These techniques can be broadly categorized into two approaches: accelerated time dynamics and direct saddle point searches. Both are designed to extract the energy barriers and reaction pathways that are essential for predicting material behavior over longer and more realistic timescales.

The class of accelerated dynamics methods aims to increase the frequency of rare events within an MD framework, thereby extending the effective simulation time. Temperature-accelerated dynamics (TAD) achieves this by running simulations at an artificially high temperature, where transitions occur more frequently, and then the dynamics is systematically extrapolated back to the desired physical temperature (e.g., Figure 2a). Hyperdynamics is an alternative method that adds a carefully constructed bias potential to the PEL, which effectively lowers the activation energy barriers without altering the true transition state locations. This allows the system to escape energy basins more rapidly (e.g., Figure 2b). Each of these techniques represents a significant advancement over conventional MD, providing the crucial kinetic parameters needed for higher-level models. While powerful, these acceleration schemes’ efficiencies are subjected to the lowest energy barrier present in the system. Under certain circumstance, the simulation can become “trapped” oscillating over a low-barrier pathway, preventing it from exploring higher-barrier events that may be critical for long-term evolution.

Saddle point search algorithms represent another class of methods, which are designed to directly locate the transition states between two energy minima on the PEL. The Nudged Elastic Band (NEB) algorithm is one of the most popular methods to accurately locate the saddle point in the hyper-dimensional PEL (e.g., Figure 3a). Nevertheless, the NEB method relies on the prior knowledge on the evolution “mechanism” (i.e., the pre-defined final states). Therefore, at certain simple circumstances—e.g., point defect diffusion in crystals, single dislocation glide, etc.—the NEB can work well. However, under the scenario of strong structural disorders and chemical complexities (often seen in heavily irradiated alloys), the PEL becomes significantly complex, making it impossible to guess in advance the viable evolution mechanisms. To address this challenge, proactive self-propelling PEL sampling techniques would be needed. In this regard, a widely used approach is the dimer method (e.g., Figure 3b), which calculates the lowest curvature mode of the PEL (via the diagonalization of the Hessian matrix) to guide the system from an initial state toward a saddle point without a priori knowledge of the final state. In the case of a very-high-magnitude structural disorder—e.g., metastable grain boundaries, the regions associated with irradiation-induced amorphization, etc.—the PEL becomes extremely rugged, and the number of possible saddle states and interconnecting pathways to the initial states can grow exponentially. To probe such a kind of complex PEL with more nuances features, the activation–relaxation technique (ART) is a very capable algorithm that has been demonstrated to be successful in even the completed disordered alloys [27,28,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65]. The ART begins from a relaxed, local energy minimum and applies a perturbation to move the system uphill along the PEL following the Lanczos algorithm [29] until a saddle point is identified. By introducing numerous random perturbations from the same initial state until the obtained activation energy spectrum converged, ART can capture a statistically more robust sampling of the complex PEL. Despite of these methods’ powerful capability, their reliance on the Hessian matrix diagonalization, under certain circumstances, might render the computational cost prohibitive.

The autonomous basin climbing (ABC) method is an innovative algorithm for exploring the PEL that overcomes many limitations of prior techniques. It is an activation–relaxation algorithm designed to systematically map reaction pathways and reconstruct the energy landscape without requiring a priori knowledge of the final states or reaction coordinates. More specifically, the process begins with the system relaxed in a local energy minimum. A series of repulsive, Gaussian-shaped penalty functions are then incrementally added to the PEL, forcing the system to climb out of its initial energy basin (e.g., Figure 3c). This continues until the system is pushed over a saddle point, at which point it relaxes into a new, lower-energy minimum configuration. By tracing the system’s trajectory, the energy barrier for the transition can be determined. This approach is particularly effective because it circumvents the need for computationally expensive Hessian matrix diagonalization, making it more broadly applicable to complex defect systems [30,31,68,69,70,71,72,73,74]. However, the original ABC method has a significant limitation rooted in its “one-dimensional” nature. By design, the algorithm tends to identify and follow the single pathway of least resistance—the one with the lowest activation energy barrier—while neglecting other, potentially competitive, higher-barrier pathways. For materials with complex energy landscapes, such as those with anisotropic crystal structures (e.g., hcp Zr) or the heterogeneous chemical environments found in advanced alloys, multiple competing defect mechanisms are common, making this a critical issue. To address this, the extended ABC (ABC-E) method was developed [32]. ABC-E enhances the exploration of the PEL by systematically cataloging multiple transition pathways from a single starting state. After the first (lowest-barrier) pathway is identified, the algorithm blocks it by adding a repulsive potential at its saddle point. The search is then re-initiated from the original basin, compelling the system to find the next-lowest energy pathway. This iterative process is repeated to identify multiple relevant escape routes, yielding a comprehensive set of activation barriers and transition mechanisms. This ensures a far more accurate calculation of the system’s true kinetics. Furthermore, the ABC framework can be implemented in a dynamic scheme to simulate defect evolutions and interactions under prescribed external conditions, such as a much more realistic strain rate that can be handled by conventional MD. As will be discussed below in Section 4, by coupling ABC/ABC-E searches with incremental changes in strain, it becomes possible to model dislocation–obstacle interactions over a vast range of strain rates, directly bridging the gap between atomistic simulations and experimental observations. This advanced capability establishes the ABC method as a uniquely powerful and robust tool for the predictive modeling of long-term microstructural evolution in irradiated materials. A brief comparison among these methods regarding their advantages and limitations are summarized in

Table 1. A summary of the features of various PEL-based atomistic algorithms.

Method	Primary Use Case	Key Advantage	Limitation/Scalability
Temp.-Accelerated Dynamics (TAD)	Accelerating rare events by manually increasing the temperature	Simulating long-time dynamics of a system in a single energy basin	Conceptually straightforward; extends MD time by leveraging high-T runs
Hyperdynamics	Accelerating rare events by modifying the potential energy surface	Boosts escape rates from energy wells without changing transition pathways	Requires the careful construction of a bias potential; also limited by the lowest barrier in the biased system
Nudged Elastic Band (NEB)	Finding the minimum energy path and saddle point between a known initial and the final state	Highly accurate for calculating the energy barrier of a specific, known mechanism	Requires a priori knowledge of the final state, making it unsuitable for discovering novel or complex mechanisms
Dimer Method	Finding a saddle point without knowing the final state	Proactively searches for an escape pathway from a given minimum	Relies on Hessian matrix calculations, which can be computationally expensive for large systems
Activation–Relaxation Technique (ART)	Exploring complex energy landscapes by finding multiple saddle points from a single minimum	Robustly samples the PEL to find a distribution of possible escape pathways	Can also be computationally demanding due to its reliance on the Hessian matrix
Autonomous Basin Climbing (ABC/ABC-E)	Systematically mapping reaction pathways and barriers without prior knowledge of mechanisms	Highly efficient as it avoids Hessian calculations; the ABC-E variant can find multiple competing pathways	The original ABC method tends to follow only the lowest-energy pathway, potentially missing other relevant mechanisms

.

3. Atomistic Modeling on Point Defect Behavior

3.1. SIAs and Vacancies’ Dynamics in Different Crystalline Structures

3.1.1. BCC Metals

Body-centered cubic (BCC) metals, such as iron, vanadium, tungsten, and their alloys, play crucial roles in the current and proposed structural materials for nuclear reactors. Consequently, their response to irradiation—e.g., the generation of SIAs and vacancies, as well as their migration over time—has profound implications for microstructural evolution and property degradation and hence has been a central focus of computational materials science [75,76,77,78,79].

In metals, it is known that, instead of occupying the geometric interstitial sites, the SIAs are more preferred to form small clusters such as dumbbells or crowdions along the closest packing direction (<111> for the BCC lattice) together with surrounding neighbors, to minimize the formation energy. These clusters are highly mobile but mostly constrained within one dimension along the <111> axis, which exhibits extremely low activation barriers for this 1D glide. Occasionally, the clusters could alter their orientation to another <111> direction and keep gliding fast along the new direction. Overall, the SIAs’ motion exhibits a 1D/3D mixture pattern. It is worth noting that, in addition to such a general pattern, there are more nuances associated with SIAs’ kinetics. For example, while transmission electron microscopy (TEM) observations qualitatively supported the fast 1D motion of SIA clusters, a notable discrepancy arose from experimental resistivity recovery studies. These experiments measured activation energies for post-irradiation recovery that were significantly higher than those predicted for simple 1D glide, suggesting that more stable, less mobile SIA configurations must exist. In particular, atomistic simulations have revealed the existence of metastable, three-dimensional, non-parallel configurations (NPCs) of SIA clusters as precursors to mobile dislocation loops [80]. As seen in Figure 4, the ABC method has been further employed to probe the underlying PEL and uncover the unfaulting pathways and associated governing barriers [68]. The obtained high barrier is consistent with the experimental data and explains why these sessile clusters can persist at higher temperatures. The metastability feature of certain SIA clusters endows them with the potential of acting as potent obstacles to dislocation motion and contributing significantly to mechanical performance (see Section 4 for a more detailed discussion).

On the vacancy side, clustering is the fundamental precursor to the formation of voids and macroscopic swelling. The evolution of vacancy populations is often described by a multi-stage recovery model, where the nature of Stage IV recovery in BCC iron has been a long-standing point of controversy [81,82,83]. Atomistic simulations have provided critical insights into the mechanisms governing this stage. By modeling a supersaturated system of vacancies, simulations have shown that small, mobile defects like di-vacancies and tri-vacancies diffuse through the lattice until they are absorbed by larger, immobile clusters. Furthermore, these studies have captured the Ostwald ripening mechanism at the atomistic level, where larger clusters grow at the expense of smaller, less stable ones that dissociate. A crucial finding from these works is the strong dependence of defect kinetics on the local environment; for instance, the migration barrier for a single vacancy is significantly lower in the vicinity of a larger vacancy cluster (0.44 eV) compared to the bulk value (0.63 eV) [6]. This kinetic detail is essential for creating accurate, predictive models of void nucleation and growth. By incorporating these atomistically informed mechanisms, simulations have successfully predicted a transition temperature for vacancy cluster mobility around 150 °C, in agreement with positron annihilation spectroscopy (PAS) experiments [84], thereby delineating the controversial Stage IV recovery process.

3.1.2. FCC Metals

Face-centered cubic (FCC) metals and their alloys, including nickel, copper, and austenitic steels, have also long served as model systems for investigating the fundamental mechanisms of radiation damage. In FCC metals, the most stable configuration for a single SIA is the <110> dumbbell, where two atoms share a single lattice site. Atomistic simulations have shown that these single SIAs are exceptionally mobile, even at cryogenic temperatures [85,86,87]. As irradiation proceeds, these SIAs rapidly cluster to form small, two-dimensional dislocation loops, typically on $\left\{111\right\}$ planes (Frank loops) or, upon unfaulting, perfect prismatic loops. As in BCC metals, numerous simulation studies show that these small SIA loops exhibit extremely high one-dimensional (1D) mobility. They can glide rapidly along their Burgers vector direction with a very low activation barrier, a phenomenon sometimes described by a “breathing” crowdion model [88]. This rapid 1D glide is a highly efficient mechanism for transporting interstitials over long distances, allowing them to be annihilated at sinks like grain boundaries, surfaces, or other defects. This process is fundamental to clearing defects from the crystal lattice and plays a critical role in the overall damage accumulation and evolution of the dislocation network.

In contrast to the highly mobile nature of SIAs, vacancies in FCC metals tend to form sessile, immobile clusters. While small vacancy clusters can have some mobility, the defining feature of vacancy aggregation in many irradiated FCC metals (such as copper, gold, silver, etc.) is the formation of stacking fault tetrahedra (SFTs), as illustrated in Figure 5. An SFT is a three-dimensional defect, appearing as a hollow tetrahedron bounded by stacking faults on the four $\left\{111\right\}$ planes. MD simulations of displacement cascades have been crucial in elucidating the primary formation mechanism of SFTs. They form directly and spontaneously from the collapse of the dense, vacancy-rich core of a high-energy cascade [89,90]. Within picoseconds, the disordered cluster of vacancies can rearrange into the highly stable, ordered SFT structure. While it is in general believed that, once formed, SFTs are extremely stable and immobile, some studies also show that the entropy effects could enhance the kinetics of SFTs beyond normal expectation [90]. SFTs could serve as localized sinks for other point defects and, more importantly, as strong obstacles to dislocation motion. This makes them a primary contributor to irradiation-induced hardening and the increase in yield stress observed in these materials [91,92].

3.1.3. HCP Metals

Hexagonal close-packed (HCP) metals, such as zirconium and its alloys, are widely used as the cladding materials in nuclear reactors. Unlike the highly symmetric cubic lattices of FCC and BCC metals, the HCP structure possesses a lower degree of symmetry, which gives rise to a profound anisotropy in its material properties [93]. This inherent anisotropy is a defining characteristic of HCP metals and is especially pronounced in the behavior of point defects, where migration and clustering are strongly dependent on crystallographic direction. Understanding these directional dependencies is critical for predicting the dimensional stability and mechanical performance of HCP components under irradiation.

Atomistic simulations have been indispensable for mapping this anisotropic behavior. For vacancies, diffusion does not occur isotropically but is instead governed by two distinct jump mechanisms: migration within the basal plane and migration out of the basal plane (i.e., with c-axis projection). Each pathway has a different activation energy barrier. For example, in zirconium, simulations using the ABC-E method revealed that the in-plane migration barrier is significantly lower than the out-of-plane barrier [32]. This directly results in anisotropic vacancy diffusion, with mobility being substantially faster within the basal plane than along the c-axis [94]. This directional flux of vacancies is a primary microscopic driver of irradiation growth, a unique phenomenon to anisotropic materials where a component changes shape under irradiation even at a constant volume [95].

The dynamics of SIAs in HCP metals are even more complex due to the existence of numerous stable and metastable interstitial sites (e.g., octahedral, crowdion, and tetrahedral sites, as illustrated in Figure 6a). This results in a multitude of competing migration pathways with different activation energies and dimensionalities. For instance, simulations in zirconium have identified both a fast one-dimensional (1D) glide mechanism for SIA clusters on the basal plane and more complex three-dimensional (3D) migration pathways involving jumps between different interstitial sites (e.g., Figure 6b). The dominant diffusion mechanism becomes a function of temperature (e.g., Figure 6c), as the system may favor different pathways depending on the thermal energy available to overcome specific barriers [32].

Such anisotropic features have direct consequences for the evolving microstructure and mechanical properties in HCP metals. The directional diffusion of point defects leads to the formation of specific populations of dislocation loops, which in zirconium are predominantly observed on prismatic planes. The subsequent interaction of mobile dislocations with this oriented population of defect clusters results in anisotropic irradiation growth, as illustrated in Figure 7. The strengthening effect is different for dislocations gliding on prismatic versus basal or pyramidal slip systems, which in turn influences the plastic anisotropy and ductility of the irradiated material. Therefore, accurately modeling the anisotropic nature of point defect dynamics is a critical prerequisite for developing predictive models of dimensional stability and mechanical integrity for HCP alloys in nuclear applications.

3.2. Effect of Chemical Complexity on Defect Behavior in Multi-Component Alloys

While the study of conventional metals provides a baseline for understanding defect physics, the frontiers of materials science have shifted toward multi-component systems, such as concentrated solid-solution alloys (CSAs) and high-entropy alloys (HEAs). These materials represent a paradigm shift in alloy design. Unlike traditional alloys with a single principal element, CSAs are composed of multiple elemental species in high, often near-equiatomic, concentrations. This compositional strategy produces a unique atomic-level landscape characterized by severe lattice distortion and a lack of long-range chemical order, which fundamentally alters the aspect of point defect behavior compared to simpler crystalline metals. Atomistic modeling has been an indispensable tool, providing the primary means to deconstruct this complexity and establish the physical mechanisms responsible for the remarkable properties, including enhanced radiation tolerance, that these alloys exhibit.

3.2.1. A Spectrum of Defect Energetics

The most fundamental difference between conventional metals and CSAs lies in their defect energetics. In a pure BCC, FCC, or HCP crystal, every lattice site is identical, resulting in single, well-defined formation and migration energies for vacancies and SIAs. In a CSA, however, the random arrangement of different atomic species means that the local chemical environment surrounding any given lattice site is unique. Consequently, a point defect does not have a single energy value but rather a broad spectrum or distribution of formation and migration energies (e.g., Figure 8). This concept has been conclusively demonstrated through atomistic simulations. First-principles quantum mechanical methods, such as Density Functional Theory (DFT), have been essential in quantifying these energy landscapes [12,97,98,99,100]. By calculating the defect energies across hundreds of unique local environments within a simulated supercell, these models have shown that the migration energy for a vacancy, for example, can vary significantly from one site to the next. This distribution of energy barriers is a core feature of CSAs and is the microscopic origin of many of their unique properties.

3.2.2. Sluggish Diffusion and Complex Defect Transport

The rugged and varied energy landscape in CSAs gives rise to kinetic phenomena not observed in pure metals, most notably sluggish diffusion. The wide distribution of migration barriers, containing numerous deep energy wells, effectively creates a multitude of local traps for diffusing defects, particularly vacancies. While a vacancy might occasionally find a low-energy path, its long-range movement is hindered by the high probability of encountering a high-energy barrier, leading to an average diffusion coefficient that is significantly lower than in its constituent pure metals [100]. The behavior of SIAs is more complex. Atomistic simulations show that, while their diffusion may also be slowed, the more critical effect is that their migration path becomes more tortuous and three-dimensional compared to the rapid 1D glide often seen in pure metals. This less-direct motion increases the time spent traversing the lattice, which in turn enhances the probability of encountering and annihilating a vacancy [102,103,104].

Furthermore, molecular dynamics simulations have revealed that defect transport in CSAs can be highly element-specific. In the FCC NiFe alloy, for instance, simulations show that vacancies preferentially migrate through exchanges with Fe atoms, while interstitials are transported primarily via Ni-Ni dumbbells. This chemical preference in defect–atom interactions means that the fluxes of different defects are intrinsically coupled to the fluxes of different elements, a critical insight for understanding and predicting radiation-induced segregation (RIS). Indeed, modeling has suggested that increasing the alloy’s chemical complexity can frustrate these preferential pathways, leading to a suppression of RIS and improved phase stability.

3.2.3. Enhanced Recombination and Reduced Damage Accumulation

The complex chemistry of CSAs also profoundly influences the initial stages of damage production and annealing (e.g., Figure 9). While early MD simulations suggested that the number of Frenkel pairs created in a primary cascade event was similar in CSAs and pure metals, more advanced simulations that include electronic effects have revealed a critical difference. The chemical disorder in CSAs enhances electron–phonon coupling and reduces thermal conductivity. As revealed by two-temperature MD (2T-MD) simulations, this causes the thermal spike generated by a cascade to be more confined and to cool down more slowly. This prolonged local annealing phase provides more time and opportunity for the newly created vacancies and interstitials to recombine, leading to a significant reduction in the number of surviving defects that can contribute to long-term damage accumulation.

This highly efficient in-cascade recombination, combined with sluggish point defect diffusion, is the primary reason for the superior radiation tolerance of CSAs. Large-scale MD simulations of overlapping cascades, which mimic high-fluence irradiation, have directly confirmed this outcome [105,106,107,108]. These studies consistently show that the growth of both dislocation loops and vacancy clusters is dramatically suppressed in alloys like NiCoCr and NiFe when compared to pure nickel under identical irradiation conditions. By limiting the population of mobile defects and hindering their ability to aggregate, the intrinsic chemical complexity of CSAs provides a powerful, built-in mechanism for mitigating radiation damage from the moment it is created.

4. Extended Defect Generation and Evolution Under Irradiation and Mechanical Loading

4.1. Stress Sensitivity of Irradiated Dislocation Loops

The generation of extended defects, primarily dislocation loops, is a direct consequence of displacement cascade events and represents the primary form of stable, irradiation-induced damage that governs the evolution of mechanical properties. Most early simulations were carried in a stress-free environment [109,110,111,112]. However, in an operating reactor, materials are simultaneously subjected to both intense irradiation and significant mechanical stress, creating a coupled environment where the production of these defects is profoundly altered. Understanding the synergy between irradiation and stress is critical, as materials are known to behave very differently under these concurrent conditions compared to when the stimuli are applied separately. Recent atomistic simulations have unveiled that the applied stress state does not merely change the quantity of damage produced but fundamentally alters the type of dislocation loops that are preferentially generated. For example, He et al. [113] employed a sophisticated MD framework, which coupled a two-temperature model for realistic cascade physics with a systematically controlled principal stress tensor, to map the stress sensitivity of defect production in an NiFe alloy. As illustrated in Figure 10, their work revealed a remarkable dichotomy in the response of the two most dominant types of extended defects: sessile Frank loops and glissile Shockley partials.

The formation of Frank loops, which are interstitial in nature and act as primary hardening obstacles, was found to be dictated by volumetric strain. Their production exhibited a strong and asymmetrical dependence on normal stresses: it was significantly boosted under tensile loading and suppressed under compressive loading. In stark contrast, Frank loop density was almost completely insensitive to pure shear stress. The underlying mechanism is tied to the formation energy of the interstitial plates that collapse to form these loops; this energy is directly lowered by tensile stress and raised by compressive stress, thus favoring or inhibiting their nucleation. Conversely, the generation of Shockley partials, which are glissile dislocations and can act as direct carriers of plastic deformation, was discovered to be governed by an entirely different mechanism. Their formation is controlled by the magnitude and relaxation rate of the local von Mises shear stress that develops at the interface between the hot, liquid-like cascade core and the surrounding solid matrix. Consequently, their production is significantly enhanced under an applied shear load but is far less sensitive to purely tensile or compressive stresses. The simulations demonstrated that an external shear load sustains the high local shear stresses created by the cascade for a longer duration, allowing more time for Shockley partials to nucleate and stabilize before the cascade energy dissipates.

These findings on primary defect production add a critical new dimension to the established understanding of stress effects on irradiated microstructures, which has historically focused on two other phenomena: the stress-induced preferential alignment of loops [114] and stress-biased defect absorption [115,116,117]. It has long been known that an applied stress can bias the orientation of nucleating loops to align with the stress field in a way that best accommodates the strain. Furthermore, the long-term evolution of the dislocation network is governed by the biased absorption of point defects at existing dislocations, which is the microscopic basis for irradiation creep. The work by He et al. fills a crucial knowledge gap by elucidating the physics of the initial defect generation step itself, which sets the stage for all subsequent evolution. This insight may potentially offer a novel pathway for materials engineering: by strategically tuning both irradiation and mechanical loading conditions, it may be possible to orthogonally control the populations of sessile (hardening) and glissile (plasticity) defects, thereby tailoring the mechanical response of alloys in real-time.

4.2. Dislocation–Obstacle Interaction Mechanisms at Various Thermomechanical Conditions

The macroscopic mechanical properties of irradiated metals—including hardening, ductility, and creep—are ultimately dictated by the collective behavior of dislocations moving through a complex microstructure populated with small obstacles [92,118,119,120,121,122,123,124,125,126,127,128,129,130]. A predictive understanding of material performance therefore requires mechanistic insights into these fundamental dislocation–obstacle interactions. Historically, a significant gap has existed between experimental observations at realistic temperatures and strain rates, and the results from conventional MD simulations are confined to athermal or high-rate conditions. For example, in situ TEM experiments show that irradiated SFTs in FCC metals can be completely absorbed by a moving dislocation [89]. However, conventional MD simulations on the same dislocation–SFT interaction show that the SFT remains intact after the dislocation passes by [91,131]. Similar discrepancies also exist in other alloys. For example, conventional MD simulations show the irradiated SIA cluster in HCP Zr cannot be absorbed by moving dislocation [122]; however, TEM characterization shows the clear formation of dislocation channels, indicating the swept of the irradiated defects by the moving dislocation in its slip plane [132].

This controversy was resolved by the long timescale atomistic simulation techniques discussed above in Section 2. For example, by combining ABC with the transition state theory (TST), a dynamic atomic-level simulation framework can be established to explore the dislocation–obstacle interaction across a wide spectrum of temperatures and strain rates. In the dislocation–SIA cluster interaction in HCP Zr, the ABC-TST simulations revealed that the two seemingly contradictory observations were not contradictory at all; rather, they are two distinct mechanisms that dominate under different thermomechanical conditions, as seen in Figure 11. In particular, the findings were synthesized into a temperature–strain rate mechanism map that delineates the governing physics: (i) At low temperatures and/or high strain rates, the system is dominated by the applied mechanical force, with insufficient time for thermally activated processes to occur. In this athermal-dominated regime, the dislocation shears and passes through the SIA loop, consistent with early static simulations [122]. (ii) At high temperatures and low strain rates, the system has ample time for thermal activation to mediate the interaction. This allows the atoms at the dislocation–obstacle interface to reconfigure, leading to the absorption of the SIA cluster by the dislocation and subsequent dislocation climb [133]. This mechanism explains the experimental observation of dislocation channeling. This work demonstrated that the interaction outcome is a direct result of the competition between the timescale of thermal activation and the timescale imposed by the strain rate, providing a unified framework that reconciled decades of seemingly disparate results.

The similar ABC-TST approach was also employed to unravel other complex phenomena, such as the origin of negative strain rate sensitivity (nSRS) in BCC iron associated with the dislocation–void interaction, as seen in Figure 12. More specifically, nSRS is a counter-intuitive behavior where a material’s strength decreases as the deformation rate is lowered, which can lead to plastic instabilities and flow localization. The ABC-TST simulations of a dislocation interacting with a small vacancy cluster (a nano-void) revealed a striking “V-shaped” relationship between the critical resolved shear stress (CRSS) and the applied strain rate. Above a critical strain rate of ~10⁵ s⁻¹, the material exhibited normal behavior, with the critical resolved shear stress (CRSS) increasing with higher strain rates. Below this threshold, however, the system entered the nSRS regime, where the CRSS decreases as a function of strain rates. The atomistic origin was again traced to the competition between the strain rate and thermal activation. At low strain rates, the vacancy cluster has sufficient time to thermally activate and restructure into a more compact, lower-energy configuration before the dislocation trespasses. This more stable structure presents a smaller contaminating surface area and hence weaker impedance to the dislocation, thus lowering the stress required for breakaway [134]. At high strain rates, there is no time for this relaxation; the dislocation interacts with a less-stable, higher-energy void structure or simply shears it apart, requiring a higher stress. These simulations provided the first direct, unit-process-level explanation for the microscopic origins of nSRS.

These detailed case studies exemplify a broader principle: the properties of the obstacles themselves—be they solute atoms, precipitates, or radiation-induced defect clusters—and their interaction with dislocations are not static but are dynamically influenced by the thermomechanical environment. By capturing these complex, time-dependent mechanisms, advanced atomistic simulations provide the fundamental insights needed to build more robust and physically grounded models of plasticity and failure in materials for extreme environments.

It is worth noting that, among many different types of defects, dislocations are arguably one of the most important kinds as they are the main carriers of plastic deformation [126,135,136,137]. In particular, the interactions between dislocations and obstacles (e.g., precipitates introduced into the matrix) largely dictate many critical mechanical properties such as precipitation hardening, strain rate sensitivity, formability, etc. [115,138,139,140,141,142,143]. Dislocation-mediated microstructural evolution is an immensely complicated phenomenon involving multiple physics over many orders of magnitude in both spatial and temporal scales. Therefore, to enable a reliable and robust multiscale modeling of such problem, it is crucial to transfer the obtained atomistic mechanisms, accurately and completely, to higher-level models.

The diagonal streamline on the left of Figure 13 represents a widely adopted multiscale modeling scheme on the dislocation–centric microstructural evolution. At the macroscopic scale, the finite element method (FEM) has been regarded as one of the most appealing and effective techniques [144,145,146,147]. The fidelity of FEM largely depends on whether its mesh can be appropriately handled, and whether there are reliable constitutive descriptions for all the nodes. The constitutive equations for dislocation-mediated deformation are usually provided by mesoscale models, such as line tension analysis, dislocation dynamics (DD), and the visco-plastic self-consistent (VPSC) method [43,148,149,150,151,152,153,154,155,156,157,158,159,160,161]. On the other hand, these mesoscale models rely on many fitting parameters or empirical assumptions, such as the dislocation core structures, the critical resolved shear stress for each slip system, the strain hardening exponent, the dislocation jog/junction formation mechanisms, the dislocation climb rate, etc. To reduce the number of empirical parameters, one still needs to employ more fundamental yet computationally more expensive modeling from atomic levels, such as the first-principles calculation and MD method [92,122,124,125,129,130,162,163,164,165,166,167,168,169]. In sum, through the hereby described paradigm, it is expected that the correct physical processes at the atomic scale can be carried and transferred into the macroscopic description.

To ensure that fundamental physics is retained, the atomic-to-mesoscale interface is of crucial importance. In most existing practices, such an interface has been built in a deterministic manner, while ignoring the fact that the same dislocation–obstacle pair might take qualitatively different interaction pathways and mechanisms at various conditions. Such treatments, in turn, weaken the true physics coupling between different processes. For example, in the current multiscale modeling of precipitation hardening, the precipitates grow by the deterministic classical kinetics theory [170,171], and their geometrical information (e.g., concentrations and sizes) are then taken to a separate and independent subroutine to uniquely determine system’s macroscopic strength following the Orowan bowing or Scattergood–Bacon relation [172,173,174,175], assuming the precipitates are impenetrable. However, should the dislocation–obstacle interaction adopt additional mechanisms other than Orowan looping—as shown above in Figure 11 and Figure 12—then the extra pathways would in turn yield different growth kinetics of precipitates, which might subsequently lead to very distinct overall microstructural evolutions. Such a scenario possibility, however, has been precluded from the current multiscale modeling paradigm.

Although first principles and MD studies have yielded notable advances in the knowledge of dislocation–mediated mechanics from the fundamental levels, they face formidable challenges in probing the non-equilibrium phenomena at realistic timescales. Specifically, typical mechanical test experiments are operated under strain rate conditions of 10⁰ s⁻¹ or lower [132,141,176]. However, regular MD studies of dislocation–obstacle interactions are performed either at extremely high strain rates (>10⁶ s⁻¹) [123,124,125,130,177,178,179] or at static conditions (i.e., T = 0 K) [121,122,170,180,181] that effectively give rise to equivalent outcomes as in the high strain rate limits due to their similarities in suppressing thermal activations [69,133,134,182,183]. It has been demonstrated that the unrealistically short timescale in regular MD simulations not only could induce huge errors in yield strength quantifications but also, more critically, could provide misleading microstructural evolution mechanisms. A remarkable controversy between the MD prediction and experimental observation, as briefly mentioned above in the beginning of Section 4.2, is illustrated in the right panel of Figure 13, where in situ TEM measurements clearly show that a perfect SFT is fully absorbed and removed by a single moving dislocation [89]. However, parallel MD simulations show that the same type of obstacle remains intact after cutting by a moving dislocation [91,92,131]. Many other materials also exhibit the similar phenomena [118,133,134], namely, the same dislocation–obstacle pair might interact in qualitatively different manners at various conditions, and mechanisms revealed by regular atomistic modeling are not always supported by real experiments. In other words, the intrinsic timescale limitation in traditional MD imposes significant challenges in discovering the complete and relevant defect interaction mechanisms governing the material degradations under real operations. Therefore, such potentially ill-informed atomistic mechanisms, together with the above-mentioned deterministic interface, could lead to high consequential loss to the fidelity of multiscale modeling on defect microstructural evolution. In such a context, given the unique advantage of PEL-centric atomistic algorithms in probing the multiple competing defect interaction mechanisms (as seen in the examples above in Figure 11 and Figure 12), they can serve as a pivot to make a detour to circumvent the flaws in the current multiscale modeling framework, as marked by the new route (green panel) in Figure 13.

4.3. Grain Boundaries Under Extreme Environments

Under the extreme conditions of severe irradiation and mechanical stress, the interfaces within crystalline materials, particularly grain boundaries (GBs), can be driven into far-from-equilibrium states (e.g., Figure 14). These non-equilibrium GBs often develop a local atomic structure and chemical composition that are strikingly analogous to those of amorphous solids or metallic glasses. This parallel is not merely qualitative; it provides a powerful, quantitative framework for understanding the mechanical behavior and failure mechanisms of irradiated materials, bridging the conceptual gap between crystalline defect physics and glass mechanics.

Atomistic simulations have been instrumental in revealing the formation process of these glassy-like interfacial layers. High-dose irradiation or severe plastic deformation introduces a supersaturation of point defects and drives significant solute segregation to GBs [11,34,117,184,185,186,187,188,189,190,191,192,193]. MD and Monte Carlo simulations have shown that this combined influx of structural damage and chemical impurities can destabilize the ordered, ground-state GB core. The GB instead reconstructs into a thickened, structurally disordered film, which can be viewed as a metastable, two-dimensional amorphous phase, sometimes referred to as a grain boundary complexion [194,195,196,197,198,199,200,201,202]. This irradiation-induced transformation effectively embeds a thin layer of metallic glass between two crystalline grains, fundamentally altering the material’s mechanical response. The most critical consequence of this structural transformation is a shift in the governing deformation mechanisms at the interface. While a more ordered, near-equilibrium GB interacts with dislocations through predictable absorption, transmission, or reflection events at certain sites and specific activation energies, the amorphous nature of a non-equilibrium GB precludes these discretized features. Instead, the atomic reconfiguration mechanisms and associated activation barriers are broadly and continuously distributed, similarly as those in amorphous solids [61,62,203,204,205,206,207,208,209,210,211,212,213,214,215,216,217,218]. The disordered, often solute-embrittled film acts as a preferential site for void nucleation and provides an easy path for crack propagation, a mechanism believed to be central to irradiation-assisted stress corrosion cracking and other forms of environmental embrittlement [219,220,221,222,223,224]. Recognizing that irradiated grain boundaries can behave like glassy solids is therefore essential for developing predictive models of material integrity and failure in extreme environments, offering a unified approach to understanding how both structural and chemical disorder dictate performance.

Figure 14. Extreme conditions such as ion irradiation could lead to roughening phenomenon of GBs and drive them far from equilibrium state (reprint from [225]).

Figure 14. Extreme conditions such as ion irradiation could lead to roughening phenomenon of GBs and drive them far from equilibrium state (reprint from [225]).

5. Accelerating Discovery with Data-Driven Methods and Machine Learning

While advanced atomistic modeling methods have provided unprecedented insights into the physics of irradiation damage, their high computational cost remains a significant bottleneck. Deciphering the complex processing–structure–property (PSP) relationships for radiation-tolerant materials requires exploring a vast parameter space, including alloy composition, temperature, stress state, and irradiation dose. Performing high-fidelity simulations for every conceivable condition is computationally intractable. This challenge has catalyzed the integration of data-driven techniques, particularly machine learning (ML), into the material design workflow. ML is not replacing physics-based modeling but is instead acting as a powerful amplifier, augmenting simulations at every stage to accelerate the discovery and design of new materials. This section explores three specific, high-impact examples of this synergy: the development of quantum-accurate interatomic potentials, the automated analysis of defect microstructures, and the creation of surrogate models for rapid performance prediction.

5.1. Quantum-Accurate Machine Learning Interatomic Potentials (MLIPs)

While quantum mechanical methods like Density Functional Theory (DFT) provide high-fidelity descriptions of atomic interactions, they are computationally prohibitive for the large system sizes and long timescales needed to simulate radiation damage phenomena during and after displacement cascades. Conversely, in traditional atomistic simulations, the employed empirical interatomic potentials (e.g., Embedded Atom Method (EAM)) are computationally efficient but often lack the accuracy and may suffer compromised performance on the modeling of chemically complex systems like multi-component CSAs and HEAs. Machine Learning Interatomic Potentials (MLIPs), as illustrated in Figure 15, have emerged as a revolutionary solution to this dilemma [226,227,228]. These are a new class of potentials that use flexible, non-linear ML architectures—such as Gaussian Process Regression or, more commonly, deep neural networks—to learn the potential energy surface directly from a large database of DFT calculations. During the training process, the ML model learns the intricate, many-body relationships between the local atomic environment of an atom and the resulting forces and system energy. Once trained, the MLIP can predict the forces on atoms with near-DFT accuracy but at a computational cost that is orders of magnitude lower, making it suitable for large-scale MD simulations.

The impact of MLIPs on irradiation studies is profound. They enable large-scale simulations of cascade dynamics and defect evolution in chemically complex alloys with quantum-level accuracy. This is crucial for capturing the subtle electronic effects and nuanced local interactions that govern defect energetics and sluggish diffusion in CSAs, which traditional potentials often fail to describe. By providing a robust and accurate foundation for MD simulations, MLIPs are unlocking new possibilities for the predictive modeling of primary damage production and defect transport in the next generation of radiation-tolerant alloys.

5.2. AI-Driven Automation of Defect Identification and Analysis

A major bottleneck in both large-scale simulations and modern in situ irradiation experiments is the analysis of the resulting microstructures. A single simulation can generate billions of atom positions, while an in situ transmission electron microscopy (TEM) experiment can produce thousands of video frames containing a complex, evolving population of defects. Manually identifying, classifying, and tracking every dislocation loop, stacking fault, and cluster in these massive datasets is slow, subjective, and impractical.

Machine learning, particularly computer vision algorithms based on Convolutional Neural Networks (CNNs), offers a powerful solution to this data analysis challenge. For the simulation output, ML models can be trained to perform rapid and accurate structural analysis, automatically identifying local crystal structures and classifying atoms belonging to various defect types [204,229,230]. This greatly accelerates the post-processing of simulation data. Even more transformative is the application of ML to experimental microscopy. Recent developments in machine learning-based automatic defect classification are being leveraged to enable high-throughput statistical analysis of TEM data [231]. These AI-driven tools can be trained on a combination of simulated and manually labeled experimental images to automatically detect defects, classify their type (e.g., Frank loop vs. Shockley partial), measure their size and density, and track their evolution (growth, shrinkage, or movement) frame-by-frame. This creates an unprecedented, high-throughput feedback loop between experimental observation and simulation, allowing for the direct, quantitative validation of atomistic models and accelerating the pace of scientific discovery.

5.3. Surrogate Modeling for Rapid PSP Exploration and Design

The ultimate goal of the research on alloys’ responses to irradiation is to predict the long-term performance of a material and to rationally design new alloys with superior properties. This requires understanding the complex, non-linear relationship between initial inputs (e.g., alloy composition, irradiation temperature, stress) and final outputs (e.g., void swelling rate, degree of embrittlement). Running a full, multiscale simulation to predict this outcome for every possible alloy composition and condition is computationally impossible. ML-based surrogate modeling provides a path to navigate this vast design space efficiently. A surrogate model uses an ML algorithm, such as Gaussian Process Regression or a neural network, to learn the input–output relationship from a limited number of strategically chosen, high-fidelity physics-based simulations [232,233]. Once this complex relationship is learned, the computationally cost-effective surrogate model can be used to make instantaneous predictions for any new set of input parameters within its training domain. This would allow researchers to rapidly map out the entire PSP landscape, perform sensitivity analyses to identify the most influential design parameters, and even perform inverse design, where a desired property is specified and the model suggests the optimal material composition and processing conditions. Such a data-driven approach could dramatically accelerate the material design cycle by intelligently guiding experimental and computational efforts toward the most promising candidates for radiation-tolerant materials.

6. Concluding Remarks and Future Outlook

The pursuit of next-generation nuclear energy systems, vital for a sustainable global energy future, is fundamentally a materials science endeavor. The extreme environments of advanced fission and fusion reactors present a grand challenge that has spurred remarkable innovation in both material design and the computational methods used to understand them. This review has traced the evolution of atomistic modeling, highlighting its indispensable role in moving the field from a largely empirical practice to a predictive, science-based discipline. By providing a window into the atomic-scale phenomena that govern material behavior under irradiation, these simulations have not only explained long-standing experimental observations but also have begun to guide the rational design of alloys with unprecedented radiation tolerance.

6.1. Summary of Key Advances and Insights

Radiation damage is an immensely complex and multiscale-nature problem, which spans from the picosecond creation of point defects in a displacement cascade to the decades-long evolution of a reactor component’s microstructure. We underscored the limitations of conventional experimental and computational approaches in bridging these vast gaps in length and time, motivating the need for the advanced methodologies that formed the core of this review. More specifically, while conventional MD remains essential for capturing the non-equilibrium physics of primary damage creation, its temporal limitations preclude the study of slow, thermally activated kinetic processes. We detailed the development of a suite of long-timescale methods—including accelerated dynamics techniques like TAD and hyperdynamics, and saddle point search algorithms like ART and the dimer method—that were designed to overcome this barrier by focusing on the underlying PEL. We placed particular emphasis on the autonomous basin climbing (ABC) method and its extended variant, ABC-E, as a uniquely powerful framework for systematically exploring complex energy landscapes without the prior knowledge of reaction mechanisms, a crucial advantage when modeling the intricate defect pathways in irradiated materials.

We then explored the fundamental behaviors of point defects in different crystal structures. Despite of certain commonalities, various crystal structures have their own distinctive signatures. For example, in BCC metals, small SIA clusters are not necessarily always mobile. Instead, they sometimes could possess metastable configurations that present very high unfaulting activation barriers, making them responsible for certain long timescale phenomena observed in experiments. In FCC metals, the vacancy-type clusters can form a unique morphology known as SFTs. As for HCP metals, the defining feature of anisotropy was shown to govern defect diffusion and, by extension, macroscopic phenomena like irradiation growth and anisotropic hardening. A pivotal insight across all systems is the transition from studying defects in pure metals to understanding their behavior in chemically complex concentrated solid-solution and high-entropy alloys. Atomistic modeling has been the primary tool to demonstrate how the rugged, heterogeneous energy landscapes in these alloys give rise to a spectrum of defect energetics, leading to sluggish diffusion and enhanced defect recombination—the microscopic origins of their superior radiation tolerance.

Building on this foundation, we also examined the generation and interaction of extended defects, which directly control mechanical properties. Recent simulations coupling cascade dynamics with an external stress field have revealed how mechanical loads fundamentally alter primary defect production, creating a stress-sensitive balance between sessile Frank loops (driven by volumetric strain) and glissile Shockley partials (driven by shear strain). Furthermore, the application of dynamic, long-timescale atomistic methods has resolved a longstanding discrepancy between simulation and experiment, such as the interaction between dislocations and irradiated SIA clusters in HCP Zr alloys. By establishing temperature–strain rate mechanism maps, these models have shown that the interaction outcome is a direct consequence of the competition between thermal activation and the applied strain rate. This same principle was shown to provide the atomic-level explanation for the origin of negative strain rate sensitivity (nSRS) in BCC iron. In addition, we also highlighted the growing analogy between the highly disordered, non-equilibrium grain boundaries found in irradiated materials and the physics of amorphous solids, suggesting a unified framework for understanding deformation and failure at these critical interfaces.

We further addressed the synergy between these physics-based models and emerging data-driven techniques. The development of Machine Learning Interatomic Potentials, the AI-driven automation of defect analysis from both simulation and experimental data, and the use of ML surrogate models for the rapid exploration of the design space are transforming the field. These methods are not supplanting physical models but are amplifying their power, accelerating the cycle of discovery, and paving the way for true materials by design.

6.2. Outlook and Newly Emergent Fundamental Questions

Despite the tremendous progress detailed in this review, our understanding of materials in extreme environments is far from complete. As our computational and experimental capabilities grow, we move from answering “what” happens to asking the more fundamental question of “why”. The next era of research will be defined by tackling deep, foundational challenges that lie at the intersection of physics, chemistry, and materials science. We propose the following five fundamental questions as guideposts for the community over the coming decades.

1. How do non-adiabatic electronic effects govern defect dynamics and energy dissipation? The current state-of-the-art simulations, including the two-temperature model (2T-MD), treat the coupling between the atomic and electronic subsystems via a phenomenological heat transfer framework. While this captures the essential energy exchange, it largely neglects the nuanced, non-adiabatic electronic effects that can occur during the violent, sub-picosecond timescale of a displacement cascade. How do localized electronic excitations, charge transfer events, and modifications to the local electronic structure influence bond breaking, defect formation, and immediate recombination? In chemically complex alloys with strong electron–phonon coupling and varying electronegativity among constituent elements, these effects could be paramount. Answering this question will require moving beyond classical potentials and phenomenological models to employ large-scale, time-dependent quantum mechanical simulations, pushing the frontiers of computational physics to directly model the entangled dynamics of electrons and atoms in a highly non-equilibrium state.

2. Can we develop a universal, predictive theory of disorder for radiation tolerance? We have established that structural and chemical disorder, as exemplified by CSAs and HEAs, is a key principle for designing radiation-tolerant materials. However, “disorder” remains a largely qualitative concept [234,235,236,237,238]. Is there a universal, quantitative metric for disorder that can predict radiation performance across vastly different material classes, from crystalline CSAs to amorphous alloys and disordered grain boundary films? Such a theory would need to move beyond simple configurational entropy to capture the statistical topology of the potential energy landscape—its ruggedness, the distribution of basin depths, and the connectivity of pathways. Developing this framework would unify the physics of irradiated crystals and glassy solids and would transform alloy design from a search within specific compositional spaces to a search for an optimal, quantifiable degree and character of disorder.

3. What are the governing principles of interfacial evolution far from equilibrium? Interfaces, particularly grain boundaries, are the primary sinks for defects and solutes, and their evolution dictates macroscopic failure. We know they transform into disordered, glassy-like films, but our understanding of their dynamic behavior is in its infancy. How do interfaces evolve—structurally, chemically, and mechanically—under the simultaneous, coupled driving forces of defect fluxes, solute segregation, mechanical stress, and chemical potential gradients? Are there universal kinetic pathways for the formation and thickening of these amorphous intergranular films? Answering this requires a non-equilibrium thermodynamic and kinetic framework for interfaces, one that treats them not as static sinks but as dynamic, evolving phases. This knowledge is essential for predicting and controlling intergranular failure modes like IASCC and high-temperature embrittlement.

4. How does the coupling of extreme driving forces alter the fundamental nature of defects themselves? Our current models largely assume that the intrinsic properties of defects (e.g., their core structure and mobility) are constant, with the external environment only biasing their behavior. But under the extreme conditions inside a future reactor, could the driving forces be strong enough to fundamentally alter the defects themselves? For example, could immense local stresses within a cascade warp the core structure of a dislocation loop, or could the extreme electronic excitation change its stable configuration? This line of inquiry challenges our basic assumptions and pushes us to consider that, in far-from-equilibrium environments, the distinction between a defect and its environment blurs. Unraveling this requires simulations where the defect is not just a participant but an emergent entity whose very nature is dictated by the extreme conditions imposed upon it.

5. Can we achieve true inverse design through an autonomous, closed-loop discovery framework? The integration of ML has dramatically accelerated material discovery, but the process is still largely human-guided. The ultimate goal is a fully autonomous, closed-loop platform that can design, synthesize, and test new materials with minimal human intervention. Can we create a system that intelligently combines physics-based simulations (powered by MLIPs) with robotic experiments (guided by AI-driven analysis) and surrogate modeling to perform true inverse design? Such a framework would allow scientists to specify a set of desired properties—for example, low swelling, high fracture toughness at 800 °C, and resistance to corrosion—and have the autonomous system discover novel alloy compositions and processing pathways that meet those targets. Achieving this vision requires not only advances in each individual component but also their seamless integration, representing a grand challenge that could revolutionize the very nature of materials science research and deliver the advanced materials needed for a clean energy future.