Kinetic Analysis of pH Effect on the Paracetamol Degradation by an Ozonation–Blast Furnace Slags Coupled System by Neural Network Approximation

Abstract

The presence of paracetamol (PCT) in aquatic environments has raised environmental concerns due to its incomplete removal in conventional wastewater treatment plants. This study evaluates the degradation kinetics of PCT using an ozonation system enhanced with blast furnace slags (BFSs) as a heterogeneous catalyst under acidic (pH 3), neutral (pH 7), and basic (pH 10) conditions. Experimental results show that a simple ozonation process achieves up to 85% PCT removal within 30 min, with the highest rates being observed at pH 10. The addition of BFSs increases the reaction rate constants by 20–30% across all pH levels, attributed to the catalytic activity of metallic oxides in BFSs, which promote radical-based degradation pathways. Biochemical oxygen demand (BOD5) and HPLC analyses confirm a significant reduction in PCT and its byproducts, while ozone consumption is optimized in the catalytic system. A hybrid kinetic modeling approach, integrating pseudo-first-order kinetics and a long short-term memory (LSTM) neural network, was developed and validated, demonstrating superior predictive accuracy (R² > 0.98) for PCT degradation dynamics compared with traditional models.

1. Introduction

Paracetamol, also known as acetaminophen (PCT, C8H9NO2), is one of the most widely used analgesic and antipyretic medications globally. This pharmaceutical product is commonly prescribed for the treatment of fever and mild to moderate pain. Its popularity surged during the COVID-19 pandemic, leading to a substantial increase in consumption worldwide [1]. In 2021, the global paracetamol market size was valued at USD 1.56 billion and it is projected to grow at a compound annual growth rate (CAGR) of 4.4%, reaching nearly USD 2.20 billion by 2029 [2]. This growth reflects the increasing prevalence of chronic diseases, flu, and viral infections that drive demand for paracetamol-based medications in both prescription and over-the-counter (OTC) forms.

The production of azo dyes and photographic materials is partially responsible for the widespread presence of paracetamol and its breakdown product (4-aminophenol) in the environment [3]. Through effluents, waste disposal from the pharmaceutical industry, and human and veterinary use, it is constantly released into the environment as a parent chemical and in the form of its metabolic derivatives [4,5]. Non-target animals found in aquatic habitats may experience harmful phenomena due to the steadily rising quantities of paracetamol and other newly developed pharmaceutical pollutants. According to a recent study, everyone should prioritize evaluating the unusually high dispersion of paracetamol in the environment and its ecotoxicological effects. Paracetamol has been identified as an emerging contaminant in aquatic ecosystems worldwide. A global study analyzed 258 rivers in 104 countries and found that one in four rivers contains residues of medications, including paracetamol, metformin, carbamazepine, and caffeine [6]. Regions such as Asia-Pacific, North America, and Europe represent significant markets for paracetamol due to rising healthcare expenditures, regulatory approvals, and the adoption of innovative drug formulations like extended-release variants and combination therapies [7].

Contamination by paracetamol and other pharmaceuticals in rivers poses risks to aquatic biodiversity and can contribute to antimicrobial resistance, potentially affecting human health. Furthermore, improper medication use, such as flushing leftovers down the toilet, exacerbates this environmental problem.

In Mexico, PCT is considered an emerging pollutant due to its inadequate disposal in wastewater treatment plants (WWTPs), resulting in the contamination of water bodies. In urban wastewater from Cuernavaca, Morelos, PCT concentrations have been found to be between 354 and 4460 ng/L [8]. In the Mezquital Valley, Hidalgo (a place in Mexico with high levels of population poverty), a study identified the presence of PCT among other pharmaceutical compounds [9], with concentrations of 78.2–112 ug/L [10]. In the Atoyac River in Puebla, contamination by severe drugs, including paracetamol, was documented, although the specific concentration was not detailed [11]. Additionally, effluents from hospitals are also a significant source of paracetamol due to the high (and sometimes uncontrolled) use of the drug in clinical facilities. These effluents are often not adequately treated before being released into the environment, with concentrations that can reach up to 38,000 ng/L, according to the studies presented [12,13]. The presence of PCT in aquatic ecosystems in Mexico represents a significant environmental issue, where it is toxic to several aquatic organisms, including microalgae, crustaceans, and fish [14]. Paracetamol has been associated with genetic variations, changes in diet, and alterations in the gonads of bivalves [15,16]; it can cause malformations in fish embryos and larvae, affecting their normal development, and this includes physical deformations and growth problems [17,18].

To mitigate the environmental impact of paracetamol, various wastewater treatment technologies have been investigated. Ozonation has emerged as an effective and environmentally friendly method for degrading organic pollutants. Conventional ozonation can reduce the concentration of diverse complex and toxic compounds; however, it may be insufficient for the complete degradation of recalcitrant substances like paracetamol [11,19]. In recent years, the modification of ozonation systems with the inclusion of some reaction conditions has led to the so-called advanced oxidation process (AOPs) [20,21]. The inclusion of hydroxyl ions, using metallic oxides, metals, or composite compounds, can produce radicals that can react with specific sections of recalcitrant compounds more efficiently than the pure ozonation system. Such a technique is commonly called catalytic ozonation [22,23,24].

The kinetic analysis of catalytic ozonation is complex and sometimes imprecise given the large number of series and parallel reactions occurring all the time that can hardly be modeled. This fact has motivated the application of complementary modeling methodologies such as those from machine learning, especially those named artificial neural networks (ANNs) [25]. ANNs are computational models that can be used to model complex chemical reactions like ozonation [26]. Because ozonation reactions involve multiple reaction pathways and are influenced by various operational parameters, ANN models provide a helpful way to determine outcomes without requiring explicit mathematical equations [27]. This fact motivates the application of ANN as a complementary model for performing more precise kinetic analysis of complex reactions such as catalytic ozonation. Among the existing forms of ANNs, long short-term memories (LSTMs) represent the state-of-the-art implementation of approximate mathematical models for systems whose dynamics involve many variables that interact in multiple ways. LSTMs contain all the processing elements that can be used to determine mathematical models of systems such as catalytic ozonation [28]. Long- and short-time gates can also help reproduce the fast decomposition of compounds during the first time, while long-term gates can reproduce the ozonation effect on the obtained byproducts.

This paper introduces a novel and sustainable catalytic ozonation system that effectively operates across a broad pH range—acidic, neutral, and basic—by utilizing blast furnace slags (BFSs), a byproduct of the steel industry, as a multifunctional catalyst for the degradation of paracetamol (PCT). The unique chemical composition of BFS, which includes CaO, MgO, Al₂O₃, SiO₂, and FeO, enables it to serve as a complex catalyst that significantly enhances the ozonation process [24,29,30,31,32,33,34]. The system’s degradation efficiency was thoroughly validated through biochemical oxygen demand (BOD5), high-performance liquid chromatography (HPLC), and ozone consumption analyses, confirming its effectiveness. A key innovation of this study is the development of a hybrid mathematical model that integrates classical ozonation kinetics with a long short-term memory (LSTM) recurrent neural network. This model was trained and tested using experimental data obtained under various pH conditions, providing a robust and accurate predictive tool for PCT degradation dynamics. Overall, this work contributes to the field by demonstrating the valorization of industrial waste as an effective catalyst in catalytic ozonation, characterizing the reaction rate constants under different pH conditions, and introducing an advanced hybrid modeling approach that combines traditional kinetics with machine learning to better understand and predict catalytic ozonation processes.

2. Materials and Methods

2.1. Chemical Reagents

Paracetamol (Tylenol, C8H9NO2, medical grade) was acquired from Jhonson & Jhonson Laboratories S.A. de C.V. 2024, Mexico City, México, in a presentation of 500 mg per pill (medical grade). The concentration of the PCT solution was 100 mg/L.

Table 1. Studying PCT systems for ozonation with and without BFSs.

System	Initial pH
PCT–SOP	3
PCT–SOP	7
PCT–SOP	10
PCT–BFS	3
PCT–BFS	7
PCT–BFS	10

shows the systems studied in this study, divided into two groups, with the first using a simple ozonation process (SOP) and the second using ozonation with BFSs, a variant of catalytic ozonation. Moreover, this study considered three different pHs, 3, 7, and 10, to study the effect of these experimental conditions on PCT degradation. The adjustment of initial pH was conducted using HCl 1.0 M (Meyer 38% purity) and NaOH 1.0 M (Meyer 97% purity). The volume of PCT solution was 0.1 L for all of the systems studied.

2.2. Blast Furnace Slags

The blast furnace slags were obtained from a steel producing factory in Puebla, México. According to its certified laboratories, the chemical components of BFS are metallic oxides, such as FeO, CaO, and SiO₂ (

Table 2. Physical and chemical characteristics of BFSs used in catalytic ozonation.

Physic	Chemical
Solid	FeO: 40%
Color grey	CaO: 22%
Density: 1.67 kg/m3	SiO2: 14%
Hardness: 7 in Mohs scale	MgO: 8.5%
Particle size: 3/8–¾ in	Al2O3: 5.3%
	MnO2: 1.6%

). These compounds could act as a common catalyst in the ozonation process to eliminate complex organic pollutants as medical compounds [24,35,36,37].

2.3. Ozonation Procedure

A laboratory-scale ozonation process was proposed in this study (Figure 1). All experiments were conducted at a controlled temperature of 25 ± 3 °C using a 250 mL semi-continuous reactor made of borosilicate glass, equipped with a porous ceramic diffuser positioned at the bottom. The initial ozone concentration of 35 mg/L was generated using a corona discharge-type ozone generator (HTU500G model from AZCO Industries Limited, Langley, BC, Canada), with an oxygen flow rate of 0.25 L/min. In the systems where blast furnace slags (BFSs) were used as catalysts, the borosilicate glass reactor packed with BFSs was irradiated with a fluorescent diode lamp emitting light at a wavelength range of 395–405 nm.

The continuous monitoring of the inlet and outlet gas streams was carried out using an ozone analyzer model BMT-930, BMT Messtechnik GmbH, Stahnsdorf, Germany. The ozone concentration sensor was connected to a personal computer, where data were collected at 0.0, 3.0, 7.0, 12.0, 15.0, 20.0, and 30 min. A dedicated data acquisition software (Matlab R2020a-9.8 version, MathWorks Inc., Natick, MA, USA) was custom designed to capture the ozone concentration variations over time, commonly called an ozonogram. Once the ozonation reaction with BFSs was finished, the slags were removed from the reactor at the top for further reutilization in the ozonation process.

2.4. Analytic Method Analysis

The samples obtained during ozonation process with and without BFSs were analyzed by UV–Vis spectroscopy to determine the PCT decomposition dynamics at 246 nm. Biochemical oxygen demand analysis at day five (DBO5) was carried out with the 890 BOD5, standard method 5210D, reactive Hach BODTrakTM II Hach, Hach Co., Loveland, CO, USA. An HPLC analysis was carried out to identify and quantify the final compounds obtained from PCT decomposition. The HPLC instrument used was the YL9100 system, UYL Instruments, Co., Ltd., Anyang, Republic of Korea. The device with solvent was degassed and used a UV detector at 210 nm. The column used was a 4.6 × 150 mm Agilent Eclipse XDB-C18 column, Agilent Technologies, Santa Clara, CA, USA, and the mobile phase was a mixture of phosphate buffer, pH = 4/acetonitrile (9:1) in isocratic elution, with a flow of 1.1 mL min⁻¹ at room temperature (acetonitrile 99.99% purity and sodium phosphate 98% purity, J.T. Baker, Phillipsburg, NJ, USA) [24].

2.5. Hybrid Modeling of Ozonation Dynamics Aided by LSTM Approximation

One helpful technique that may offer additional knowledge of the water treatment process based on the oxidant power of ozone and its derivatives is the mathematical modeling of ozonation. Given the number of series and parallel reactions that occur during the breakdown of complex compounds like PCT, kinetic modeling of ozonation is a challenging task. In this investigation, it is considered that the reactor initially contained just one contaminant (PCT). Then, regarding ozone and contaminants, the reaction between the two is called a second-order irreversible reaction.

$$\begin{array}{c}PCT+z_i{O}_3\to B_i\\ -r_D=k_{PCT}{C}_{PCT}{C}_{O_3}\end{array}$$

where $C_{PCT}$ is the concentration of PCT and $B_i$ represents each byproduct formed by the reaction of ozone with the main contaminant. The element $z_i$ is the stoichiometric ratio between ozone and the PCT. The reaction rate of the $PCT$ decomposition corresponds to $r_D$. If $C_{O_3,in}$ is the ozone concentration at the input of the reactor, $C_{O_3,r}$ is the ozone concentration for a given time consumed by the reaction with the PCT. Because all the experiments were executed in a batch regimen, there is not $C_{DEX,in}$. If the assumption is that ozone is in excess in the reactor, then

$$C_{PCT}={-C}_{PCT,r}$$

(1)

$$C_{O_3}=C_{O_3,in}-C_{O_3,r}-C_{O_3,out}$$

(2)

$$C_{O_3,r}=z{C}_{PCT,r}$$

(3)

where $C_{PCT,r}$ and $C_{O_3,r}$ are the reaction concentrations of PCT and ozone. $C_{O_3,out}$ represents the concentration at the output of the reactor, which is determined by the mass transfer between the liquid phase and the head space of the reactor. The reaction balance implies that

$$C_{O_3}=m_{PCT}-C_{O_3,out}$$

(4)

where ${m_{PCT}=C}_{O_3,in}-z{C}_{PCT,r}$. This model only uses the reaction between ozone and the contaminant or its byproducts.

It is generally accepted that the described modeling strategy gathers the contribution of ozone over the reaction kinetics throughout the reaction rate, inducing the analysis known as a pseudo-first-order ozonation reaction. This simplified technique uses the same method described in this section but characterizes the contribution of ozone in the reaction rate $k_{PCT}$, and their equivalents are considered in the reaction of the byproducts formed and decomposed during the ozonation as well as the accumulated ozonation byproducts. Nevertheless, the pseudo-first-order model may not aggregate all the kinetic variations that affect the temporal evolution of PCT. Hence, the application of a complementary data-driven component such as a neural network can help to reduce the lack of certainty on the relationship between the PCT dynamics with ozone and potential catalytic components that may appear by the interaction between ozone and the metallic oxides on the slag’s particles. Among the potential data-driven alternatives, recurrent neural networks (RNNs) could reproduce complex dynamics that may consider feedback information, such as in the cases of kinetic equations describing an ozonation reaction.

One kind of recurrent neural network that can pick up dependency in sequence prediction issues is the long short-term memory (LSTM) network. The output from the previous step serves as the input for the next RNN phase. The issue of RNN long-term dependence was addressed, whereby the RNN can estimate the evolution of a dynamic system based on the current input but is unable to characterize the variables tendency stored in long-term memory. The performance of the RNN becomes inefficient as the gap length rises. The LSTM may store data for a very long period by default. Time-series data processing, prediction, and categorization are all performed using it. Then, based on the superior characteristics for the dynamic identification of LSTM, consider the following kinetic model (in its discretized version based on the forward Euler algorithm) that preserves the interaction between the measurable components in the reaction (contaminants, byproducts, and ozone):

$$C_{PCT}\left((k+1)T\right)-C_{PCT}\left(kT\right)=-T{k}_{PCT}{C}_{PCT}\left(kT\right)C_{O_3}\left(kT\right)+LSTM\left(C_{PCT}\left(kT\right)C_{O_3}\left(kT\right)\right|\theta )$$

(5)

here, $k_{PCT}$ is the reaction rate constantly associated with the PCT variation, and $C_{PCT}\left(t\right)$ is the time variation of the main contaminant during ozonation. The term $LSTM\left(C_{PCT}\left(kT\right)C_{O_3}\left(kT\right)|\theta \right)$ represents the dynamics of the LSTM (including its free parameters $\theta$), which is given by

$$LSTM\left(C_{PCT}\left(kT\right)C_{O_3}\left(kT\right)|\theta \right)={\sigma }_c\left(W_{fc}{\sigma }_c\left(h\left(kT\right)\right)+b_{fc}\right)$$

(6)

where

$$h\left(kT\right)=o\left(kT\right)\odot {\sigma }_2\left(c\left(kT\right)\right)$$

(7)

$$o\left(kT\right)={\sigma }_g\left(W_o{x}_{\kappa }+R_o{h}_{\kappa -1}+b_o\right)$$

(8)

$$c\left(kT\right)=f\left(kT\right)\odot c\left((k-1)T\right)+i\left(kT\right)\odot g\left(kT\right)$$

(9)

$$f\left(kT\right)={\sigma }_g\left(W_f{\left[C_{PCT}\left(kT\right),C_{O_3}\left(kT\right)\right]}^T+R_{f}h\left((k-1)T\right)+b_f\right)$$

(10)

$$i\left(kT\right)={\sigma }_g\left(W_i{\left[C_{PCT}\left(kT\right),C_{O_3}\left(kT\right)\right]}^T+R_{i}h\left((k-1)T\right)+b_i\right)$$

(11)

$$g\left(kT\right)={\sigma }_c\left(W_g{\left[C_{PCT}\left(kT\right),C_{O_3}\left(kT\right)\right]}^T+R_{g}h\left((k-1)T\right)+b_g\right)$$

(12)

where $W_{i,f,g,o}\in {\mathbb{R}}^{n\times m}$, $R_{i,f,g,o}\in {\mathbb{R}}^{n\times n}$, $W_{fc}, b_{i,f,b,o}\in {\mathbb{R}}^n$, and $b_{fc}\in \mathbb{R}$, $n$ is a parameter to define the order of the function as an approximation, that is, the precision level for the approximation. $\theta$ is the parameter that represents the matrices $W_{i,f,g,o}, R_{i,f,g,o}$, the vectors $b_{i,f,b,o}$, $W_{fc}$, and the constant $b_{fc}$. ${\theta }^{\ast }$ is the parameter that optimizes the function. ${\sigma }_g\left(y\right)={\left(1-e^{-y}\right)}^{-1},{\sigma }_c\left(y\right)=\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{h}\left(y\right)$, and $\odot$ is the Hadamard product.

This simple is proposed to represent a pseudo-monomolecular reaction that describes the decomposition of the PCT. Based on model 1, and using the experimental information of $C_{PCT}\left(kT\right),C_{O_3}\left(kT\right)$, the following representation can be obtained

$$\begin{array}{c}y\left(kT\right)=\mathrm{F}\left(C_{PCT}\left(kT\right)C_{O_3}\left(kT\right)|\overline{\theta }\right)+\epsilon \left(kT\right)\mathrm{F}\left(C_{PCT}\left(kT\right)C_{O_3}\left(kT\right)|\overline{\theta }\right)\\ =-T{k}_{PCT}{C}_{PCT}\left(kT\right)C_{O_3}\left(kT\right)+LSTM\left(C_{PCT}\left(kT\right)C_{O_3}\left(kT\right)\right|\theta )\\ y\left(kT\right)=C_{PCT}\left((k+1)T\right)-C_{PCT}\left(kT\right)\\ \overline{\theta }={\left[{k_{PCT},\theta }^T\right]}^T\end{array}$$

(13)

where the variable $y\left(tkT\right)$ is the approximation of the time variation of the variable under analysis, i.e., the evolution of PCT concentration. The term $\epsilon \left(kT\right)$ describes the approximation error produced by the implementation of the differentiator. The parametric identification problem presented in Equation (13) can be solved by the well-known least mean squares method. This method cannot be applied directly because the signals of $C_{PCT}\left(kT\right)$ and $C_{O_3}\left(kT\right)$ are measured with different sampling times. Therefore, an interpolation algorithm was applied to homogenize the number of samples that can be used in the parametric identification method using an equal time span with similar sampling time.

The interpolation algorithm used an approximation based on third-order polynomials (cubic) of the ozonogram. The model proposed in (9) includes that vector $\overline{\theta }$ represents the unknown parameters in the model, and the function $\mathrm{F}\left(C_{PCT}\left(kT\right)C_{O_3}\left(kT\right)|\overline{\theta }\right)$ collects the experimental data in the model. The time $kT$ represents the sampling time ($1\le k\le K$). The term $\epsilon$ represents the modelling error obtained as the solution of the least mean square model. The solution of the parametric modelling strategy is

$$\widehat{\theta }\left(k+1\right)-\widehat{\theta }\left(k\right)=-{\alpha }_k{\nabla }_{\widehat{\theta }}J\left(\widehat{\theta }\right)$$

(14)

$$J\left(\widehat{\theta }\left(k\right)\right)={\displaystyle \frac{1}{2}}\sum _{s=1}^k{\left(y\left(kT\right)-\mathrm{F}\left(C_{PCT}\left(kT\right)C_{O_3}\left(kT\right)|\overline{\theta }\right)\right)}^2$$

(15)

3. Results and Discussion

3.1. Ozone Consumption

Figure 2 presents the ozonograms of all the systems studied. The dynamics can be divided into two groups: one corresponding to the single ozonation processes (SOPs) and the other involving ozone combined with BFSs (OP/BFS) at three different pH values (3, 7, and 10). In all cases, the initial average ozone concentration starts at approximately 35 mg/L and decreases rapidly within the first 5 min of the reaction. This behavior represents the initial stage of PCT decomposition driven by the action of the oxidant reagents present in the reaction system. SOP systems exhibit a more pronounced minimum ozone concentration, particularly at pH 3, suggesting greater reactivity and higher ozone consumption due to a direct ozone–PCT reaction mechanism.

In contrast, ozone/BFS systems show less marked decreases in ozone concentration, indicating moderated ozone consumption and PCT decomposition. This modulation is attributed to molecular interactions rather than radical-based mechanisms. The observed ozonogram dynamics highlight behavior dependent on both pH and system composition. At acidic pH, higher ozone consumption suggests increased reactivity. Meanwhile, in systems containing BFSs, the reaction is modulated, suggesting an alternative mechanism involving molecular interactions rather than free radicals.

3.2. Paracetamol Decomposition Dynamics and pH Effect

Figure 3 illustrates the effect of pH on PCT decomposition. In SOP systems, the decomposition of PCT is faster and more efficient, achieving near-zero PCT concentrations in less than 10 min. This phenomenon aligns with the corresponding ozonogram for each selected experimental condition, where a more pronounced initial ozone consumption reflects the more considerable reactivity of the oxidant reagents involved in the specific reaction. In contrast, in OP/BFS systems, PCT decomposition occurs more slowly compared with SOP systems. This outcome is attributed to the modulation of ozone consumption by the presence of certain oxides in the BFSs, which stimulates non-radical mechanisms.

A joint analysis of Figure 2 and Figure 3 provides insights into how experimental conditions influence the dynamics of ozone consumption and PCT degradation. In SOP systems, rapid ozone consumption drives efficient PCT decomposition, particularly under acidic conditions. Conversely, in OP/BFS systems, slags modulate these processes, promoting stabler but slower mechanisms.

Table 3. Initial and final pHs and electrical conductivities obtained under the selected experimental conditions.

Reaction System	pHi	pHf	Electrical Conductivity (µS/cm)i	Electrical Conductivity (µS/cm)f
PCT–SOP pH 3	3	3	54.9	413
PCT–SOP pH 7	7	3.5	77.7	319
PCT–SOP pH 10	10	3.6	106.8	127.7
PCT OP/BFS pH 3	3	6.9	201.1	226.2
PCT OP/BFS pH 7	7	6.7	80.8	239.6
PCT OP/BFS pH 10	10	7.2	108.4	208.1

reinforces the observations made from Figure 2 and Figure 3, highlighting key differences between SOP and OP/BFS systems. SOP systems are highly reactive, generating acidic and ironic byproducts that promote rapid PCT decomposition. In contrast, OP/BFS systems exhibit less dynamic behavior, with a more buffered chemical environment due to the stabilizing effect of slags. This reduces the accumulation of ionic byproducts and results in a neutral pH. Electrical conductivity increases significantly in SOP systems compared with OP/BFS. This increase is associated with the formation of ions derived from PCT decomposition, such as carboxylates or NOx, which elevates the concentration of charged species in the solution.

The results of this study demonstrate the effectiveness of ozonation, both as an SOP and in the presence of BFSs, for PCT degradation under different pH conditions. The dynamics presented in Figure 2 and Figure 3 show that the SOP achieves 65% of PCT removal at pH 7 after 30 min and up to 85% at pH 10 after 30 min. These results are consistent with the fact that SOP is more effective to pollutants degradation at basic conditions; in this case, this pH dependence aligns with previous findings, suggesting increased ozone decomposition and radical formation at higher pH levels, thereby enhancing the oxidative degradation of organic contaminants. Specifically, at higher pH, ozone can react to form hydroxyl radicals (-OH), which are powerful, non-selective oxidants.

The incorporation of BFS as a catalyst resulted in a marked increase in the reaction rate constant across all pH levels. The enhanced performance is attributed to the catalytic properties of BFSs, including metallic oxides (CaO, MgO, Al₂O₃, SiO₂, and FeO), which facilitate hydroxyl radical generation and promote alternative degradation pathways [18,29,30,31]. The experimental data indicate that the catalytic ozonation system achieved up to 30% faster PCT removal compared with conventional ozonation, especially under basic conditions, confirming the synergistic effect between ozone and BFSs. This improvement can be explained by considering that the composition of the BFSs, specifically the presence of FeO, Al₂O₃, and SiO₂, facilitates surface reactions that both decompose ozone into reactive oxygen species and directly oxidize PCT, as described in Table 2 of this study.

3.3. Effect of Fe³⁺/Fe²⁺ on PCT Decomposition

The use of visible light (395–405 nm) in the systems with BFS was intended to promote photocatalytic activity. While BFS is not a traditional photocatalyst, the irradiation could potentially excite surface electrons in the metallic oxides, leading to enhanced radical formation and improved PCT degradation. Previous studies have shown that the semiconductor properties of iron oxides, such as iron oxide, can be activated under visible light, leading to the generation of electron-hole pairs that can drive oxidation reactions [38,39]. Additionally, the synergistic effect of ozone and visible light irradiation has been observed in other advanced oxidation processes, further supporting the potential enhancement in PCT degradation. The experimental setup is described in Figure 1.

The highest metallic oxide in BFS considered in this research is iron oxide with 40% composition. Iron, in its oxidation states Fe²⁺ (ferrous) and Fe³⁺ (ferric), plays a key role as a catalyst in AOPs, especially when it is combined with ozone. This type of mechanism takes advantage of the iron’s ability to generate hydroxyl radicals (OH), which are highly reactive molecules and effective in the degradation process of organic compounds such as PCT.

According to [40,41,42,43], the radical’s formation mechanisms initiate with the ozone decomposition under visible light with λ > 380 nm:

$$O_3+H_{2}O+hv\to H_2{O}_2+O_2$$

(16)

When Fe²⁺ from BFS is present in the system, a secondary reaction was carried out to produce hydroxyl radicals [40]

$${Fe}^{2+}+O_3\to {FeO}^{2+}+O_2$$

(17)

$${FeO}^{2+}+H_{2}O\to {Fe}^{3+}+·OH+{OH}^{-}$$

(18)

The $H_2{O}_2$ obtained from ozone react with double bonds of PCT chemical structure, increasing decomposition rates. Moreover, $H_2{O}_2$ reacts with ${Fe}^{2+}$ via Fenton’s reaction, producing ·OH that also contributes with the PCT decomposition [44,45,46]:

$${Fe}^{2+}+H_2{O}_2\to {Fe}^{3+}+·OH+{OH}^{-}$$

(19)

Additionally, ${Fe}^{2+}$ can be obtained from ${Fe}^{3+}$ reduction with hydroperoxyl radical $\left({HO}_2^{·}\right)$, which is a weaker oxidant than $·OH$ [45]

$${Fe}^{3+}+H_2{O}_2\to {Fe}^{2+}+H^{+}+{HO}_2^{·}$$

(20)

$${Fe}^{3+}+{HO}_2^{·}\to {Fe}^{2+}+H^{+}+O_2$$

(21)

The presence of those oxidants has a positive effect on PCT decomposition, involving different paths: photolysis of ${Fe}^{3+}$ complexes generate carboxylic acids such as oxalic from PCT decomposition, which cannot be oxidized by ozone, hydroxyl radicals in the ${Fe}^{2+}/O_3$ systems; production of hydroxyl radicals by the oxidation of ${Fe}^{2+}$ with $H_2{O}_2$ formed by Fenton’s reactions (4) or by photoreduction of the ${Fe}^{3+}$ to ${Fe}^{2+}$ are produced by the following reaction [47]:

$${Fe}^{3+}+H_{2}O+hv\to {Fe}^{2+}+·OH+{OH}^{-}$$

(22)

Hence, the combined systems take ozone reactivity as the non-selective action of hydroxyl radicals to improve the PCT degradation. The ${Fe}^{3+}/H_2{O}_2/O_3/hv$ systems have a synergistic effect due to ozone decomposition on the BFSs, reaching a high PCT decomposition efficiency; according to [48], this suggests that the dominant mechanism is the radical-mediated degradation by the photo-Fenton reactions to increase the hydroxyl radical’s concentration (Equation (17)) [47].

3.4. Effect of MgO on PCT Decomposition

Another oxide in BFSs that plays a significant role in the PCT degradation is MgO with 8.5% composition. According to [49,50], MgO is an appropriate ozone catalyst due to its high surface activity and reactivity, high stability in water, and more importantly, less toxic and environmentally friendly properties.

A. Mashayekh-Salehi et al. [49] describe that the high catalytic activity of MgO could be explained by the interactions between ozone and the specific surface of MgO catalyst and the formation of hydroxyl radicals ($·OH)$ according to the following reaction mechanisms:

$$MgO-s+O_3\to O_2+{MgO}^{{-s}^{·O}}$$

(23)

$${MgO}^{{-s}^{·O}}+{2H}_{2}O+O_3\to {2O}_2+3·OH+{MgO-s}^{{(·OH)}_2}$$

(24)

Hence, in this case, there are several oxides that generate hydroxyl radicals that improve the PCT degradation.

3.5. Effect of Al₂O₃ and SiO₂ on PCT Decomposition

As shown in Table 2, Al₂O₃ (5.3%) and SiO₂ (14%) are the primary components of the BFSs utilized in PCT degradation. According to A. Ikhlaq et al. [50], these components do not facilitate ozone decomposition, suggesting that the PCT degradation occurs via a non-radical mechanism, in contrast to other BFS components that promote hydroxyl radical generation from ozone. The presence of Al₂O₃ and SiO₂ positively influences the system by enhancing hydrophobic and electrostatic interactions, enabling both the adsorbate and adsorbent to adopt a protonated state. Additionally, these components act as a medium to keep the contact between the BFS, ozone, and PCT. In this specific scenario, the oxidizing species are generated through molecular interactions rather than the presence of free radicals.

3.6. Final Compound Identification

Table 4. Final concentration of oxalic acid for each system.

Experimental System	Oxalic Acid Concentration (mg/L)
PCT–SOP pH 3	81.62
PCT–SOP pH 7	61.29
PCT–SOP pH 10	95.96
PCT OP/BFS pH 3	74.01
PCT OP/BFS pH 7	71.58
PCT OP/BFS pH 10	107.13

shows that the oxalic acid concentrations align with the trends observed in ozone consumption and PCT degradation. At acidic pH (3), the high ozone consumption and efficient PCT degradation in SOP systems result in higher oxalic acid concentrations. In contrast, the OP/BFS systems, with slower ozone consumption and more moderate degradation, produced lower levels of oxalic acid at pH 3, although still significant.

At neutral pH (7), the SOP systems showed lower oxalic acid concentrations, corresponding to more moderate ozone consumption and slower PCT degradation. The OP/BFS systems, on the other hand, exhibited similar levels of oxalic acid production, indicating that the presence of BFS helps to stabilize the system and sustain PCT degradation, even under less reactive conditions.

At alkaline pH (10), the higher oxalic acid concentration in the OP/BFS system (107.13 mg/L) suggests that the slags enhance the reaction under these conditions, promoting greater PCT degradation through alternative, non-radical pathways. This result further emphasizes the role of slag components in modulating the degradation process and enhancing oxalic acid production, even when ozone consumption is less pronounced.

3.7. Kinetics Study

Based on the proposed strategy,

Table 5. Reaction rate for each analyzed system.

System	pH	Reaction Rate Constant ${\mathit{k}}_{\mathit{P}\mathit{C}\mathit{T}}, \mathbf{L}/\left(\mathbf{m}\mathbf{o}\mathbf{l}\mathbf{\ast }\mathbf{s}\right)$	Confidence Interval
PCT	3	123.12	[−12.87,+12.87]
	7	247.34	[−5.61,+5.61]
	10	432.15	[−13.98,+13.98]
PCT + BFS	3	216.39	[−16.34,+16.34]
	7	329.12	[−11.45,+11.45]
	10	617.83	[−25.23,+25.23]

presents the estimated reaction rate values for PCT during conventional ozonation under the specified experimental conditions.

Notice that the significant difference between the constants k_PCT seems to imply that the decomposition of PCT produces several intermediate compounds (not identified), which can be decomposed continuously by ozone to end as OA.

Ozone consumption was more efficient in the presence of BFSs, suggesting that the catalyst optimizes ozone utilization. The kinetic reaction variations in each system (conventional ozonation vs. catalytic ozonation) reflect these underlying chemical processes. Conventional ozonation follows a pseudo-first-order model, with a rate constant that is dependent on pH. In contrast, catalytic ozonation exhibits a more complex kinetic behavior due to the interplay between ozone decomposition, radical generation, and surface reactions on the BFS catalyst. The LSTM model was particularly valuable in capturing these complex interactions.

Biochemical oxygen demand (BOD5) and HPLC analyses corroborated the efficiency of the catalytic system, showing a substantial reduction in PCT and its byproducts. The LSTM model exhibited superior predictive accuracy, capturing nonlinear interactions in the reaction system, which are not addressed by classical kinetic models.

4. Conclusions

The present study quantitatively demonstrates the effectiveness of ozonation for paracetamol (PCT) degradation across acidic, neutral, and basic pH conditions, both with and without the addition of blast furnace slags (BFSs) as a catalyst. Kinetic analysis revealed that ozonation rates are strongly influenced by pH, with acidic conditions favoring faster degradation, while the incorporation of BFSs enhanced overall degradation efficiency by 20–30%, particularly under neutral and alkaline conditions. This catalytic effect is attributed to the metallic oxides in BFSs—primarily FeO, MgO, and Al₂O₃—which promote alternative degradation pathways. The Fe²⁺/Fe³⁺ redox cycling facilitates hydroxyl radical generation, MgO enhances ozone decomposition, and Al₂O₃ and SiO₂ stabilize molecular interactions, collectively contributing to a more efficient and stable degradation process. Experimental results showed that catalytic ozonation with BFSs achieved up to 85% PCT removal within 30 min, with reaction rate constants increasing by up to 30% compared with conventional ozonation. Biochemical oxygen demand (BOD5) and HPLC analyses confirmed significant reductions in PCT and its byproducts, while ozone consumption was optimized in the catalytic system. A key innovation of this work is the development of a hybrid kinetic model combining traditional pseudo-first-order kinetics with a long short-term memory (LSTM) neural network. This model captured the complex nonlinear and temporal dynamics of the ozonation process, achieving a predictive accuracy with R² > 0.98, outperforming conventional kinetic models. The LSTM model integrated experimental variables such as ozone consumption, pH variations, and intermediate product formation, providing a robust framework for understanding and optimizing catalytic ozonation. These findings underscore the potential of BFSs as a sustainable catalyst in advanced oxidation processes and highlight the value of data-driven modeling to enhance treatment efficiency and predictability in wastewater remediation.