High‐Throughput Photocatalysis for Generating Reliable Datasets Analyzed by Machine Learning

Abstract

Photocatalysis is an environmentally conscious tool for removing contaminants from water. Novel photocatalytic materials are often measured on ability to degrade a small number of analytes, which may not be indicative of broader applicability. In this work, an experimental method dubbed high‐throughput photocatalysis (HTP) is introduced to assay photocatalytic materials against a range of analytes in a time effective manner. HTP is modular; experimental parameters, including matrix, can be changed to fit a proposed application. The photodegradation of each analyte is attained in a consistent manner such that machine learning (ML) models can be applied to the obtained datasets. Three out of the box ML models—linear regression, random forest (RF), and neural network (NN)—are tasked with estimating the percentage removal as a function of irradiation time and molecular structure, as represented by Morgan fingerprints. Leave‐out sets demonstrated that RF and NN models did not overfit the training data and reasonably estimated the degradation of unknown molecules. SHapley additive exPlanations values are utilized to correlate molecular substructures to the parent molecule's susceptibility to photocatalytic degradation. These correlations are used to generate heatmaps of estimated reactivity within molecules that corroborate reports in which dye degradation pathways were studied in detail.

Photocatalytic materials are typically tested on limited analytes, limiting real‐world relevance. High throughput‐photocatalysis, enabling rapid, modular experimentation across 27 analytes is introduced. Using P25 TiO₂, dye removal datasets with machine learning is paired. Random forest and neural networks predict degradation, while SHAP analysis highlights molecular substructures linked to susceptibility or resilience, visualized through molecular heatmaps.

1. Introduction

Heterogeneous photocatalysis is a widely studied chemical process in which semiconducting nanomaterials (photocatalysts) absorb and convert light to useful chemical energy with potential applications in areas such as water remediation^{[ 1} , ² , ^{3 ]} and as active components in self‐cleaning surfaces.^{[ 4} , ^{5 ]} The mechanism for water remediation involves the generation of electron‐hole pairs (e⁻ _CB and h⁺ _VB) in the conductance and valence bands of the photocatalyst when light of sufficient energy is absorbed. The electron and hole migrate to the surface, where they can directly reduce or oxidize organic species at the surface or generate reactive oxygen species (ROS) such as the superoxide and hydroxyl radical.^{[ 6 ]} The efficiency of a photocatalyst is typically measured by the rate or completeness of degradation of a particular organic pollutant. This efficiency is also dependent on a variety of experimental conditions, including the composition (chemical^{[ 7} , ^{8 ]} and physical^{[ 9} , ^{10 ]}) of the photocatalyst, structure of the analyte^{[ 11 ]} and the matrix^{[ 12 ]} (the solution in which the reaction is taking place). Majority of reports in the field of photocatalysis have focused on the development of novel photocatalytic materials^{[ 13} , ^{14 ]} with strong efficiency and broad applicability to target pollutants of emerging concern. However, understanding how the structure of the analyte influences its susceptibility to photocatalytic degradation can help in determining if photocatalysis is an appropriate remediation technique.

Structure‐activity relationships have been used to describe the behavior of small molecules in several chemical processes.^{[ 15} , ¹⁶ , ^{17 ]} In photocatalysis, Romao et al. provide an extensive survey of the susceptibility of pharmaceutical compounds to photolytic and photocatalytic degradation by P25 TiO₂ and UV light.^{[ 18 ]} Their report relates increased removal efficiency to decreased molecular size and conjugation, as well as the presence of particular functional groups (aldehydes, alcohols, ketones, and nitriles). Huang et al. provide a complementary approach, studying a small set of analytes, each of which is a sulphonamide‐containing antibiotic.^{[ 19 ]} This focused scope allowed them to quantitatively relate molecular properties such as the energy of the HOMO, maximal values of nucleophilic attack, and minimum negative charge to a molecule's susceptibility to photocatalytic action.

A general structure‐activity relationship for photocatalysis can help provide valuable insight into catalyst design as well as into the strengths or limitations of the technique when deployed. In recent years, the integration of high‐throughput experimentation with machine learning (ML) techniques has garnered considerable attention across scientific domains.^{[ 20} , ²¹ , ^{22 ]} This fusion presents a promising avenue for accelerating research processes by efficiently handling the vast amount of experimental data. One of the most important requirements of utilizing ML is the availability of a high‐quality dataset on which models can be trained, tested and validated. A popular strategy to obtain such datasets is through the “scraping” of literature, where the results of various research groups and studies can be amalgamated into a single dataset.^{[ 23} , ^{24 ]} This automated dataset construction can be powerful, but it relies on the consistency in the experiments and in reporting of experimental parameters and associated results in the original studies. Photocatalytic dye removal efficiency, for example, is highly sensitive to a number of experimental parameters such as light intensity,^{[ 25 ]} reaction vessel size,^{[ 26 ]} temperature,^{[ 27 ]} catalyst concentration,^{[ 28 ]} and so on. These parameters are inconsistently reported in the literature, which has made direct comparison of results from different studies difficult, and the amalgamation of historical results into a cohesive dataset practically impossible. To date, only a handful of photocatalytic datasets have been constructed through literature scraping, and those works focus on photocatalytic water splitting.^{[ 29} , ^{30 ]}

An alternative to literature scraping is generating self‐consistent datasets, which either requires long experimentation times using traditional techniques or the development of high‐throughput techniques. Bi & Westerhoff demonstrated a high‐throughput analysis of TiO₂ formulations utilizing a 96‐well plate across a small number of analytes and reaction conditions.^{[ 31 ]} Yanagiyama et al. extended this work to a broader number of analytes and environmental conditions, creating a platform for optimizing photocatalytic water purification.^{[ 32 ]}

We undertake a similar approach in this article, utilizing 96‐well plates to simultaneously conduct multiple photocatalytic reactions for efficient and consistent high‐throughput (HTP) data collection under various controlled reaction conditions. This approach produces consistent and reproducible datasets that can be used for ML to elucidate the effects of target molecules and reaction environments on speed of the photocatalytic reaction. By building upon the insights gleaned from prior studies, we aim to further elucidate the practical considerations and methodological refinements pertinent to the utilization of 96‐well plates for HTP data collection, thereby advancing our understanding of their utility in facilitating efficient and scalable experimentation across various scientific disciplines and providing steps toward a general structure‐activity relationship for photocatalysis.

2. Results and Discussion

In order to properly compare the photocatalytic degradation of various organic dyes, experiments should be done under the same conditions. In this work we use P25 as the photocatalyst, investigating its action on 27 dyes to obtain efficiency of its action under varying conditions. We designed a modular experimental framework dubbed “High‐Throughput Photocatalysis” (HTP) that allows for 14 dyes to be studied simultaneously at a time with photolytic controls (i.e., with light on but with no photocatalyst) allowing the entire experiment (27 dyes) to be evaluated in ≈3 h considerably reducing experimentation time over traditional single dye methodologies. This framework allows for the physical conditions of the reaction (catalyst, matrix, light source, etc.) to be modified, and the resulting datasets can be compared to elucidate the effect of the change on a large number of analytes. Typically, catalysts are only evaluated on a small number of analytes, and their use against a broad range of species is not properly captured. The HTP framework described below allows for catalysts to be screened on a wide range of potential pollutants and can be used to identify potential niche applications for newly developed materials.

2.1. High‐Throughput Photocatalysis

Twenty‐seven organic dyes were initially chosen for these studies. These are listed in Table  1 , and their absorbance spectra are in Figure S2, Supporting Information. These dyes were selected because they contain a representation of a breadth of functional groups while also being compatible with the experimental framework: they are sufficiently soluble and have reasonably strong molar extinction coefficients. They range in molecular weight from 200–900 g mol⁻¹, with the exception of Direct Red 80 (DR80) at 1235 g mol, Supporting Information, and are entirely composed of organic moieties except for Luxol Fast Blue MBSN (FB‐MBSN), which contains a porphyrin stabilized Cu(II). The presence of certain functional groups (azo, alcohol, ketone, and amide) within the members of this dye set are detailed in Table 1. A number of azo‐containing dyes were selected due to their prevalence in previous photocatalytic literature, as well as their relevance in industrial wastewater. The work of Romão et al. found that alcohol, ketone, and amide groups were positively correlated with high degrees of photocatalytic degradation^{[ 18 ]} and were thus given consideration here. And since formal charges on dyes have often been used to explain observed trends in photocatalytic behavior,^{[ 33 ]} cationic, anionic, zwitterionic, and neutral dyes were included. Absorbance of visible light is a requirement and limitation for degradant species studied using this technique. UV absorbance is incompatible due to the strong absorbance properties of the photocatalytic nanomaterials.

Table 1.

Structure and select moiety assignment for organic dyes utilized in this work.

Each 96‐well plate is designed to monitor the photocatalytic and photolytic discoloration of 14 dyes in triplicate, with two lanes reserved to hold a catalyst scattering blank and for a photolytic discoloration of Erythrosin B (Ery B). This lane serves as a control since Ery B is known to photolytically degrade with first order kinetics; monitoring its degradation between experiments was used to ensure that there are no significant fluctuations in light levels. A control experiment with Ery B present in each well of the plate was carried out (Figure S3, Supporting Information) to show that the use of a UV crosslinker (CL‐1000 in this work) as the light source was adequate to provide equal illumination between each well.

2.2. Photocatalytic Degradation Curves

In photocatalytic reactions, several heterogeneous and light‐dependent processes occur in parallel, which can lead to the degradation and subsequent removal of an organic dye from solution. Dye photolysis, the breakdown without an added photocatalyst, was not found to be a major source of degradation in this study, except in the case of Bromophenol Blue, which was found to undergo significant discoloration in each of the matrices studied. Dye removal can also occur via adsorption to the colloidal surface without degradation. Dye adsorption can be inferred in this experiment by comparison of dye concentration at time 0 between photocatalytic and photolytic samples (Figure  1 black and red, respectively). While it was not explicitly calculated, adsorption to the P25 was found to be minimal, consistent with the reported poor adsorptivity of organic species on TiO₂.^{[ 34 ]} Any observed discoloration of the solution can thus be attributed to photocatalytic degradation of the dye by P25.

Figure 1.

The measured time dependance of the discoloration of organic dyes in the presence of: both P25 and UV light (black lines); UV light only (red lines), carried out in different matrix conditions: RO eater (squares); 6 mM ethanol (EtOH, an oxidative interference agent; triangles); 6 mM chloroform (CHCl₃, a reductive interference agent; circles).

The degradation curves for each of the 27 dyes utilized in this study in each of the three solution matrices studied are shown in Figure 1. The dyes chosen for this dataset show a breadth of resiliency toward photocatalytic degradation. In RO water, three dyes (CB, DR80, Congo) are quite resilient, showing less than 35% removal after 15 min of irradiation, while three others (BF, Fl, and NBA) are quite susceptible, showing greater than 80% removal within that time. These results cannot be directly compared with those of other studies due to differences in experimental conditions (catalyst conc., light intensity, irradiation path length, etc.), but general trends on the effectiveness of P25 can be observed. Methylene Blue was found to degrade quicker than Methyl Orange, consistent with a previous study that reported reactivity of both dyes under the same conditions.^{[ 35 ]} The variability of P25's effectiveness against various organic dyes is expected, as photocatalysts typically outperform with certain dyes compared to others, due to dye susceptibility to produced ROS.

An unusual photocatalytic result was observed for Methyl Green, Basic Fuchsin, and Malachite Green, where the absorption peak increased at early time points. The cause of this increase is still under investigation. Each of these dyes contain a pararosaniline core which could indicate that this structure may be relevant to the observation; however, this pattern did not extend to Crystal Violet, another pararosaniline derivative. The increase in absorbance was observed in the photolysis samples and could appear to hinder degradation rate at low irradiation times as calculated using (Equation (S1), Supporting Information).

Introducing radical interference agents (also called “scavengers” in earlier literature) to the reaction matrix and monitoring the effects on the degradation efficiency is a common means of probing the degradation pathway.^{[ 36 ]} EtOH was used as an oxidative interference agent and CHCl₃ as a reductive interference agent; these were separately introduced into the solution at 100x the initial dye concentration so that any photogenerated ROS will preferentially interact with the interfering agent instead of the dye. Hydroxyl radicals (.OH) generated from holes in the valence band of P25 are often seen as the primary agent responsible for its photocatalytic activity.^{[ 37 ]} All dyes in this work except for Reso and NBA are consistent with that view, showing a decreased in removal percentage when ethanol is present. Further support for these results is the observation of these dye's increased removal when in the presence of CHCl₃, which interferes with the reductive species that is expected to impede the rate. Nile Blue A was found to degrade at the same rate in the presence of and absence of ethanol, suggesting that competition for photogenerated oxidative species is not rate‐limiting in its case. While a small enhancement in the presence of CHCl₃ is observed for Nile Blue A, which suggests that an oxidative pathway may still be dominant, the observed effect is muted in comparison with the other dyes studied in this work. Observations into apparent degradation pathways is another application of the screening potential of this framework, as many organic molecules can be screened for uncommon catalyst—dye interactions that warrant future investigation.

It was observed that dye ‘families’—that is, those with similar structures – appear to be affected in similar manners across each of the matrices. Two examples of dye families are the rhodamine dyes R6G and RhB, and the Coomassie dyes BBR and BBG. The degradation of the rhodamine dyes was relatively quick, with 50% degraded after 15 min. This degradation was strongly attenuated in the presence of EtOH and enhanced in the presence of CHCl₃. The Coomassie dyes, on the other hand, degraded to a similar amount in RO water. Their degradation levels were enhanced in the presence of CHCl₃, while not being strongly influenced by EtOH.

The removal percentage was not found to be strongly related to dye charge (Figure S4, Supporting Information), which suggests that charge alone was not a significant predictor of degradation. We use this observation, together with the established structural consistencies, to serve as a basis for the development of a model structure‐activity relationship. The goal of this model is to be able to correlate the structural components that result in a greater or lesser degree of degradation by the catalyst in the matrix of interest. The analysis of dataset of this article, with 27 dyes that have a diversity of functional groups, lead to a large combinatorial complexity, which is not tractable in a simple way, like traditional curve‐fitting methodology.

2.3. Catalyst Reproducibility

The experimental design for high‐throughput photocatalysis must be capable of generating high quality and reproducible datasets that can be suitable for analysis with ML models. One factor that can affect this reproducibility is the consistency of the catalyst: these are particulates that have a range of sizes and shapes and may have fluctuations even within specific lots. In order to investigate reproducibility within the commercial catalyst lot, two separate 0.3 mg mL⁻¹ stock P25 solutions were prepared, and the full experimental compliment was run as described in methods section (Supporting Information). Figure  2A shows the degradation rates obtained by running each experiment plotted against one another. If both stock solutions showed equivalent performance at degrading dyes, each point would be located on the 1:1 dotted line through the center while if one stock outperformed the other for a particular dye, the resulting data point would deviate from the reference line (two red dotted lines are provided as reference for a doubling of degradation rate). Stock 1 outperformed Stock 2 considerably, with several dyes degrading at double the rate despite being prepared at the same concentration. This has significant implications for ML as it would introduce an unacceptable amount of variance into the dataset for any model training.

Figure 2.

Reproducibility study of HTP methodology. The HTP experiment is conducted twice in each figure, with A) being conducted from separately prepared P25 0.3 mg mL⁻¹ stocks at the same concentration and B) being conducted from the same P25 0.3 mg mL⁻¹ stock. The rates of both trials are plotted against each other. Dyes which appear on the 1:1 line have the same rates between trials, with 2:1 or 1:2 dotted lines representing a doubling or halving. RO was used as the matrix.

This result can be directly contrasted with the variability between dilutions done from a particular 0.3 mg mL⁻¹ stock solution. Figure 2B shows observed rate of degradation for replicate 0.05 mg mL⁻¹ dilutions from a singular stock. The rates, when plotted between replicates, fall closer to the 1:1 rate line suggesting that the catalytic performance does not vary significantly between replicates. These tests were performed one month apart, suggesting stock solutions of catalysts can be stored for this period of time. The error bars used in Figure 2 are the standard deviation of the observed rate constant (Equation (S2), Supporting Information) across three triplicate wells per experimental plate. These error bars are not sufficient to fully account for the differences in observed rate between replicates (i.e., deviation from the 1:1 line), suggesting some persistent variability.

The above results indicate that conducting experiments from one catalyst stock solution reduces the experimental variance significantly and gives a more accurate measure of the true reproducibility of the experiment. For this work, results are only considered internally comparable if they are generated from the same 0.3 mg mL⁻¹ stock solution. While this restriction currently limits the ability to compare different catalysts, it provides a higher degree of confidence in the differences observed between matrices.

2.4. ML Models

ML is utilized in this work to analyze the observed results in order to gain insight into which functional moieties contribute to a dye's susceptibility or resilience to photocatalytic degradation. The models developed, as discussed below, were not optimized to be fully predictive for unknown species, nor to converge to low error with minimal samples. Since the goal of this work is not (yet) to produce a general treatment of photocatalysis, the photodegradation results obtained in the various solution matrices are kept as three separate datasets. Models are trained on each dataset independently, and the results between datasets are treated qualitatively.

Dye structures are represented using extended connectivity fingerprints (ECFP), which is an algorithmic process specifically designed to identify molecular substructures for use in developing structure‐activity relationships.^{[ 38 ]} For this study, they were implemented using the GetFingerprint function in RDKit using default invariants. This process utilizes the Morgan algorithm to produce fingerprints similar to the ECFP family. A fingerprint, or substructure, is a computational representation of a portion of a molecule centered around a particular atom (underlined below). This can include individual atoms (e.g., C, N, O, etc.), which are said to have a radius of 0, and simple (C‐NH₂, C‐OH, N=N‐C (radius of 1)) or complex functional groups (e.g., C‐C=C‐C‐N= which represents an aromatic carbon meta to an azo nitrogen). The combination of a unique set of fingerprints can be used to reconstruct a molecular structure. Molecules with similar structures will have similar fingerprints, as much of their backbone will be represented by the same substructures but will differ by fingerprints corresponding to their unique chemical features. The Morgan algorithm will take inputs of radius and bit array length and generate potential substructures, which are then used to represent molecular structures of dyes present in the dataset. A radius of 2 (allowing substructures to include atoms 2 units away from the central atom) was chosen to provide increased differentiation between organic dyes with similar core structures. 4096 was chosen to be the length of the bit array used to store substructures. This number was increased from typical numbers such as 1024 to minimize collisions wherein two different substructures would be assigned to the same position but not increased further to avoid overfitting and reduce the computational demands. These parameters generated 407 unique substructures and 11 bit collisions present in the set of organic dyes studied. The number of unique substructures and frequency of collisions as a function of bit vector length are shown in Figure S11, Supporting Information. The 407 nonzero bits represent the molecular substructures generated by the Morgan algorithm which were present in dyes in this dataset. 3689 substructures generated (nearly 90%) were not present in the dataset, and although passed into the analysis models discussed below, have no impact as they are 0 or ‘off’ for all samples. Methyl Orange contained the least nonzero bits at 27, while Brilliant Blue G contained the most with 79, with the average dye of the set containing ≈45 nonzero bits. Despite efforts to minimize collision frequency, some were still observed, such as bit 202, which represents a ∝‐C to an alcohol in certain molecules but a β‐C to a nitrogen atom in others. Every atom is not necessarily described by an individual bit; however, all atoms are represented via the associated radius of a bit describing an adjacent atom (i.e., Table 1A. MB sulphur). A similarity matrix of fingerprints produced by the methods above was calculated using Tanimoto Coefficients and is shown in Figure S13, Supporting Information. The only highly correlated (>0.8) pair of dyes was Crocein Scarlet 3B and Orange G while an additional 19 pairs showed correlation >0.4.

Three ML models, a linear regression (LR), a random forest (RF), and a neural network (NN), were exposed to the obtained datasets and tasked with estimating the removal percentage as a function of the irradiation time and ECFP fingerprint. To accomplish this, each sample in the dataset is ascribed a target parameter (the removal percentage) and descriptors (ECFP fingerprint and irradiation time). A sample in this work constitutes the removal percentage of a dye after a single set irradiation time. Irradiation times of zero, where removal percentage is zero, were not included in the training or validation sets. The LR, RF, and NN models were selected for study in this work due to their established efficiency and differing architectures. LR provides a simple, interpretable model based on the assumption of linear relationships between features and the target variable. RFs improve prediction accuracy by averaging multiple decision trees, effectively handling complex feature interactions. NNs are well‐suited for modeling nonlinear relationships, leveraging multiple layers to capture intricate patterns, though they require more computational resources.^{[ 39} , ^{40 ]}

In order to evaluate model performance, the R‐squared coefficient of determination (R²) and mean squared error (MSE) were calculated for the trained model on both training and test sets. Each model produced reasonably high R² (>0.9) and low MSE (<0.04) on the training set for RO water (Figure  3 ), as well as the training sets for EtOH and CHCl₃ (Figure S5 and S6, Supporting Information, respectively). Both LR and NN maintain similar performance on the testing set while RF performs noticeably worse under both metrics. High marks in both metrics show that the dataset fits nicely with the modeling process as well as each algorithm can identify the patterns largely well. The steady performance among models manifests the stability and trustworthiness of the database. The poor scores of RF on the test set is conserved while analyzing the EtOH and CHCl₃ datasets as well (Figure S5,S6, Supporting Information). Model learning curves generated through cross validation can be used to gauge a model's fit of the data (Figure S7, Supporting Information). RF and NN models show similar learning curves, neither indicating significant over‐ or under‐fitting of data. The NN cross‐validation score asymptotically approaches the training score with diminishing error, which suggests a well‐fit model. RF similarly increases but does not achieve the same R² score as NN and maintains significantly larger error. LR's performance on training data, which decreases with increased cross validation, indicates that LR is not an appropriate model for this dataset.

Figure 3.

Summary of mean squared error (MSE) and correlation (R²) performance metrics for three trained models (linear regression (LR), random forest (RF), neural network (NN)) on photocatalytic degradation dataset in RO water.

2.5. “Leave‐out” Sets

In order to gauge the resiliency of the models, an analysis was performed using the data set while omitting the results of a particular dye from the training set. This process evaluates the model's ability to predict using a fingerprint it did not see during the training process and can serve as another benchmark in model performance. For this discussion, Amaranth (Am) was randomly selected as the subject of the “leave out” set, but its results were typical when other dyes were excluded instead. The leave out sets for all other dyes in the study can be found in Figure S12, Supporting Information and it can be observed that the models both over and under estimated dye degradation while also predicting different orders for the rates of degradation. This suggests that while a limited number of dyes were utilized in training, they are diverse enough in structure to not produce significant overfitting.

Figure  4 shows a summary of results when the three models are applied to the same dataset (RO data), excluding Am from the training set and including it in the testing set. RF and NN models provided a similar MSE and R² score compared to the full dataset, while LR shows a significant increase in MSE and a large reduction in R² in comparison. The consistent performance among models demonstrates the stability and reliability of the database, as well as the models’ ability to accommodate the significant requirement for generalized predictions, even when specific dyes are removed. Each trained model was then used to predict the degradation reaction of Am for comparison with experimental results.

Figure 4.

Summary of model performance metrics (MSE and R²) for LR, RF, and NN models on the RO water dataset when a representative dye, Am, is removed from the training set and reintroduced into the testing set. Note that testing set results for LR are beyond the axis limits.

As Table  2 shows, LR is incapable of generating a realistic degradation curve and predicts that 1.7 × 10¹¹% of the dye would be removed after 25 min of irradiation; clearly this is not a realistic outcome. LR's prediction inadequacies occur when a fingerprint bit had not been exposed to a model during training. As Am contains at least one unique bit that occurs only in its fingerprint (as do all other dyes in the dataset), the coefficient that the model assigns to said unique bit has not been tuned for and therefore generates a nonsensical output. In general, the model coefficients of the descriptors generated by LR are incredibly large values (often 10⁹–10¹¹) suggesting that the model is highly‐tuned and is over‐fitting the dataset. Additionally, overfitting to the training data can make the model sensitive to new data patterns, and the LR assumption of linear relationships may fail if the actual relationship is more complex. It is only because multiple iterations of each dye are present within the full dataset (with varying degradation times) and that the test/train split consistently represents each dye at least once that LR is able to achieve reasonable benchmarks using the metrics outlined above. Despite this, the overfitting and complexity of the relationship between dye structure and degradation rate demonstrate that LR should not be used to predict degradation of species not present in the training data.

Table 2.

Comparison of model‐predicted removal percentage in RO water to the actual removal percentage for a specific selected dye: Am.

Model	Linear regression [LR]	Random forest [RF]	Neural network [NN]
Predicted removal [%]	1.7 × 1011	70	72
Experimental removal [%]	86	86	86

RF and NN both predict a total degradation around 70% after 25 min of irradiation, significantly lower than the experimental value of 86% for Amaranth. Both models show realistic degradation curves that underestimate experimental values by 16%. This discrepancy is within the acceptable level of tolerance for predictive modeling errors and suggests that even while the models could not fit removal percentage perfectly, they can still provide an adequate predictive power. This is further evidenced when examining model predictions when other dyes in the dataset were chosen as the “leave out”. Predictions of “leave out” set dye were split between under‐ (Figure  5a Am), over‐ (Figure 5b BBR), and proper (Figure 5c BBG) estimations when compared to experimental results. These deviations from experimental results arise from inherent complexities in the photocatalytic process. The ability of the models to adequately predict degradation within this margin shows their overall effectiveness in modeling the degradation.

Figure 5.

Compares the predicted RO degradation curve with the actual degradation curve for specific selected dyes: a) Am, b) BBR, and c) BBG.

2.6. Model Interpretability Using SHapley Additive exPlanations (SHAP) Values

SHAP values are a common means of interpreting ML models^{[ 41 ]} and can be applied to each of the models studied in this work. A SHAP value is associated with each feature present in the dataset, including reaction time and each of the structural bits, and explains how the value of each feature affects the model's prediction for a particular sample.

Larger SHAP values indicate that a particular feature or bit is more impactful when the model makes a prediction. By analyzing which bits have large SHAP values, insight into which substructural features the model identifies as important to predicting degradation can be gleaned. Figure  6 shows the top 20 features by SHAP value for LR, RF, and NN models for the RO Dataset (SHAP values produced by each model for other datasets are presented in Figures S8–S10, Supporting Information). Positive SHAP values indicate that the presence of a particular bit increases the predicted degradation, while negative values suggest a decrease. The blue and red dots on the central plot represent the SHAP values for individual samples, with blue indicating low feature values and red indicating high feature values. The color gradient bar shows the feature value scale from low (blue) to high (red). In this dataset, the molecular fingerprints are treated as binary with a 1 (high:red) value if a substructure is present regardless of the number of instances a substructure is present, and a 0 (low:blue) otherwise. The quantification and ranking of SHAP values helps in identifying which molecular fragments are most influential in the model's predictions, providing insights into structure‐activity relationships in the dataset.

Figure 6.

SHAP values for the 20 most important features for each class of trained model (top: LR, middle: RF, bottom: NN). The chemical structure of each bit that appears in the 10 most important features is shown. For RF and NN, the top 20 features and their SHAP values are obtained from 50 trained models (5 test/train splits with 10 starting seeds). For LR, the top 20 features and their SHAP values are aggregated across 50 unique test/train splits.

In general, chloride ions, oxygen atoms, nitrogen atoms, aromatic carbons, and carbons adjacent to nitrogen functional groups have a significant impact on the model's output, despite each model ranking the bit positions differently. Most bits in the top 10 of all models are functional groups or the atom next to the functional group of the molecules. Functional groups often contain atoms with high electronegativity differences, leading to polarized bonds. This electron distribution creates regions of partial positive or negative charge, attracting nucleophiles (electron‐rich species) or electrophiles (electron‐poor species), respectively. Additionally, functional groups can participate in resonance, which delocalizes electrons over several atoms. This delocalization stabilizes reaction intermediates and transition states, making reactions more favorable at these sites. For example, carbonyl groups can stabilize adjacent carbocations through resonance, enhancing reactivity and facilitating the occurrence of reactions at these locations.

NN models outperformed LR and RF in both performance metrics and predictive capabilities, implying SHAP values obtained from trained NN models are better interpreters of the model's insight compared with the other two models. Bit 202 (a conjugated carbon with an OH group) has a high impact on the model's prediction of removal percentage, influencing the output on both the positive and negative sides. Bit 1866 (a chloride ion, Cl‐) significantly increases the model's prediction of removal percentage at higher values. The heteroatom functional groups Bit 3200 (an amine group, R₂NH) and Bit 1917 (a carbonyl group) both have a negative impact on the model output with high feature values. Conjugated carbon fragments (Bit 1391 and Bit 1088) show both high and low feature values, but the impact on the model is very low. However, when the conjugated carbon fragments are near heteroatoms such as nitrogen (Bit 191 and Bit 2891), they show a significant negative impact on the model for some data points. Hydroxyl groups (Bits 3650 and 807) show a more mixed impact, with Bit 3650 tending to increase the removal percentage at higher values, while Bit 807 tends to decrease it.

2.7. Molecular Heatmaps

Molecular heatmaps can be generated to visualize the SHAP values for different atoms within dye structures (Figure  7 , showing results for MO, R6G, and MB) across three predictive models. Atomic color are assigned based on the position of the SHAP value relative to the entire dataset. Red regions within a heatmap signify that a submolecular fragment is often present in molecules with lower‐than‐average degradation rates, while green signifies the opposite. The physical meaning of a large SHAP value cannot be directly interpreted by this analysis, only the model's relative inferred importance of the submolecular fragment on the discoloration process. Large SHAP values could be inferred as the fragment's importance for structural stability, wherein the fragment is targeted by photogenerated species, or the fragment's importance to the conjugated system responsible for the parent molecule's visible absorption, which could be disrupted by photocatalytic reactions.

Figure 7.

Molecular heatmaps of select organic dyes showing regions of molecules that correspond to higher (green) or lower (red) estimated degradation across trained models (Top: LR, Middle: RF, Bottom: NN). Selected dyes were chosen as their degradation spans the average of the dataset, with MO degrading slower than average, R6G showing average degradation, and MB degrading faster than average. SHAP values used to generate heatmaps are aggregates of 50 trained models (LR: 50 unique test/train splits; RF and NN: 5 Test/Train splits with 10 starting seeds).

The three dyes shown in Figure 1 were chosen as they represent structures with varied removal behavior by P25 (the reaction speed for MO is slower than average, R6G is around average, and MB is faster). As the SHAP values are relative to the entire dataset, molecules with quicker removal will have a greater density of atoms with positive (green) values while resilient molecules will appear more red. Generally, heteroatoms in the conjugated fragments are favorable reaction spots, such as the azo group in MO, the sulpur atom in MB, and the amine groups in all three dyes, which the SHAP bit values highlight. Additionally, conjugated carbon fragments, especially those near other carbon rings or heteroatoms, are also reaction‐favored spots, as reflected in the SHAP bit values.

Across all heatmaps, RF, and NN models exhibit similar reaction hotspots for the three dyes, indicating consistent patterns in their predictions. However, different spots on the heatmaps show varying degrees of reaction correlations between descriptors and removal percentage, reflecting the models’ sensitivity to different features. In contrast, the LR model displays significant differences compared to RF and NN, particularly for some atoms (i.e., the aromatic carbons present in MB) where it shows opposite correlations. This divergence of results between LR and RF and NN suggests that the simpler linear approach of LR fails to capture the complex relationships present in the data. Given the results of model evaluations discussed above, it is evident that LR is not suitable for this dataset.

For MO, the removal percentage is lower than average, resulting in a predominantly red heatmap with varying shades of color. Darker red shades indicate that the NN and RF models’ predicted reactions at these locations on the dye are more resilient to degradation in relation to the entire dataset. The models reveal that the azo group and the conjugated carbons near the azo group exhibit deep red colors, indicating that these structures in the dye are not favored for degradation reactions. This finding is consistent with the literature, which proposes that the first step of MO's degradation involves breaking the bonds in the amine group and sulfate group.^{[ 42 ]} This consistency between the model predictions and the literature supports the validity of the models’ findings.

For R6G, the removal percentage is around the average value, resulting in both red and green colors appearing in the heatmap. Different models assign varying, and sometimes opposite, weights to certain atoms. The models suggest that molecules containing oxygen on the ring and conjugated charged amine groups are favored for the degradation reaction, while secondary amine functional groups are less favorable. However, literature indicates that the initial step of R6G degradation involves the two amine groups, with the oxygen atom not participating until the very last step.^{[ 43 ]} This discrepancy between the models’ predictions and the literature suggests that the models may not fully capture the reaction mechanisms of R6G, potentially highlighting the need for further refinement or additional data to improve model accuracy.

For MB, the removal percentage is above the average value, resulting in a predominantly green heatmap with varying shades of color. Darker green shades indicate that the models predict that reactions at these spots are relatively more significant across the entire dataset. The models show that the chloride counterion, the sulfur atom on the ring, and the conjugated carbon near the nitrogen on the ring exhibit deep green colors, which indicate that these structures in the dye are highly favored for degradation reactions.

As ECFP was utilized to generate submolecular fingerprints, the tertiary amines of MB are treated as distinct features. Notably, these groups are assigned opposite contributions to the degradation reaction. Literature on the degradation mechanism of MB shows that the first step involves bond breaking or oxidation of heteroatoms on the aromatic ring, specifically the sulfur and nitrogen atoms, which supports the models’ findings.^{[ 44 ]} However, the literature also suggests that tertiary amine groups are not involved in the early stages of the degradation reaction. This difference indicates that the models cannot correctly handle and recognize the same structures within the molecule, highlighting a limitation that may require further refinement for more accurate predictions.

2.8. Matrix Effects on Model Predictions

The use of radical interference agents such as ethanol and chloroform can alter the dominant photodegradation pathway by reducing the active concentration of certain ROS. By analyzing how matrices that contain interference agents affect the model's prediction, some insight into degradation pathways can be gleaned. The analysis below was performed using a NN model due to its performance on the RO dataset. The EtOH dataset displays a slight drop in testing R² relative to the others but still maintains a good score, which is evidence for its reliability as a model for both degradation prediction and analysis (Figure  8 ).

Figure 8.

Summary of model performance metrics (MSE and R²) for NN models on three separate datasets of photocatalysis carried out in: RO water; with added 3 mM EtOH, an oxidative interference agent; and with added 3 mM CHCl₃ a reductive interference agent).

As discussed in the photocatlaytic results above, the presence of the oxidative interference agent ethanol significantly slowed the removal rate for the majority of dyes, except for Reso and NBA), which suggests an oxidative species was responsible for the dominant discoloration pathway. Figure  9 shows the molecular heatmaps generated for MO, NBA, and MB by the NN model in the presence of the two interference agents.

Figure 9.

Molecular heatmaps generated for three representative dyes (MO, NBA, and MB (left to right)) by trained NN models with datasets obtained in the presence of different radical interference agents. Each heatmap is the aggregate SHAP value generated by 50 NN models (5 test/train splits with 10 starting seeds) trained independently on each dataset.

For the degradation of MO in the presence of ethanol, the azo and sulfate groups lose their red coloration (Figure 9 ‐ middle) and return to a neutral white. This suggests that the NN model finds these functional groups less resilient to discoloration in the presence of ethanol, and a reductive attack on the molecule involving these functional groups may be the dominant pathway. For results with CHCl₃, the methyl groups become increasingly associated with rapid degradation. As the presence of CHCl₃ would accelerate oxidative reactions, this further supports the results of the RO example where this is the assumed path.

NBA is an interesting case in this dataset as it is the only dye not significantly affected by the presence of either EtOH or chloroform, which suggests that its removal is not hindered when either the oxidative or reductive pathways are limited. The heatmaps generated by the NN model agree with this finding, with minimal change in apparent importance between each matrix.

For MB in the presence of ethanol, the amine and conjugated carbon become a deeper red, indicating their resiliency to discoloration is maintained in the reductive environment. The thiazine nitrogen becomes significantly more green suggesting that molecules with this functional group may be targeted by reductive pathways. In the presence of CHCl₃, the entire molecule has green coloration, further confirming the oxidative pathway presented in previous works.^{[ 39} , ⁴⁵ , ⁴⁶ , ^{47 ]}

3. Conclusion

In this work we have shown an approach for generating an internally comparable dataset that enables comparison of photocatalytic reactions. The degradation of twenty‐seven organic dyes with a single photocatalyst was investigated. The standardized framework presented here allows for changes in the environmental matrix as well as other conditions for the photocatalytic reactions. By doing all experiments under the same controlled conditions, good‐quality data sets were obtained that can feed into advanced analysis, such as by ML approaches. By carrying out experiments in parallel, acquisition of data took a reasonably short fraction of the time than typically required in photocatalytic studies. The datasets generated in RO water and with small concentrations of ethanol and chloroform were used to train and evaluate LR, RF, and NN models. NN models were found to produce correlative and predictive results with the dataset attained through this work. SHAP values extracted from trained models were used to interpret the results and quantified the importance of various molecular substructures in relation to the degradation reaction. Analyzing changes in model prediction provide deeper insights into the photocatalytic degradation processes, showing how particular molecular substructures are more favored or resilient in oxidative or reductive environments. This technique provides an experimental platform which can be utilized to screen novel photocatalytic materials against an array of organic contaminates, investigate pathways of degradation, and rapidly obtain data while changing experimental parameters.

Supporting Information

The authors have cited additional references within the Supporting Information.^{[ 30} , ^{31 ]} The experimental details and further model performance information can be found withing the Supporting Information.

Conflict of Interest

The authors declare no conflict of interest.

Author Contributions

Mark Croxall: conceptualization (lead); data curation (equal); formal analysis (equal); investigation (lead); methodology (lead); writing—original draft (equal); writing—review and editing (equal). Reece Lawrence: conceptualization (supporting); data curation (equal); formal analysis (equal); investigation (equal); methodology (equal); visualization (equal); writing—original draft (equal); writing—review and editing (supporting). Jiaqi Gong: conceptualization (supporting); data curation (equal); formal analysis (equal); investigation (equal); methodology (equal); software (lead); visualization (equal); writing—original draft (equal); writing—review and editing (equal). M. Cynthia Goh: funding acquisition (lead); resources (lead); supervision (lead); writing—original draft (equal); writing—review and editing (equal). Mark Croxall, Reece Lawrence and Jiaqi Gong contributed equally to this work.

Supporting information

Supplementary Material

Croxall Mark, Lawrence Reece, Gong Jiaqi, Cynthia M., ChemPhysChem, 2025, 26, e202500039. 10.1002/cphc.202500039

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request. All raw data and code used for analysis are availible at: https://github.com/Goh‐Lab/Datasets‐for‐TiO2‐Generate‐by‐HTP.git.