MatSci-ML Studio: an interactive workflow toolkit for automated machine learning in materials science

Design philosophy and scope

To realize the objective of creating an accessible yet powerful ML platform, MatSci-ML Studio was designed based on a clear philosophy. The primary goal is to empower materials scientists and domain experts who have limited programming backgrounds by automating the most laborious and technically demanding stages of the data-driven workflow.

The toolkit is specifically engineered to handle structured, tabular datasets, such as those used for modeling composition-process-property relationships. MatSci-ML Studio is not intended as a universal solution for all data types; it does not, for instance, natively process unstructured data such as microstructural images or spectra without prior, user-led feature extraction. Its core strength lies in providing a robust, end-to-end pipeline for the rich information contained within well-structured tables.

The integrated workflow architecture

The software’s modular architecture, presented in Figure 1, logically encapsulates each stage of the ML pipeline into a series of interconnected, user-friendly tabs. Built using PyQt5^[11], the GUI systematically guides researchers through the entire discovery process. The core system is organized around the following key modules, each designed to address a specific challenge in the materials informatics workflow.

Figure 1. Flowchart for the MatSci-ML Studio. ML: Machine learning.

Project management and collaboration

A core pillar of MatSci-ML Studio is its robust project management system, which is designed to ensure methodological rigor and enhance reproducibility. The module facilitates the creation of dedicated projects, each organized into a structured directory that isolates data, models, results, and configuration metadata. Building on this foundation, MatSci-ML Studio incorporates a critical version control feature, enabling users to create timestamped “snapshots” of the entire project state. This capability captures the exact data, preprocessing steps, and model parameters utilized at any given point. This allows researchers to readily revert to previous stages or compare different experimental workflows, thereby guaranteeing full traceability.

Export and Share: Projects can be exported as a single, self-contained archive, which simplifies sharing of complete workflows and findings with collaborators.

Data management and quality assessment

The workflow commences with the data management module, which provides intuitive tools for data ingestion and initial exploration.

Flexible Data Import: The module supports loading data from various common formats, including CSV, Excel (.xlsx, .xls), and directly from the system clipboard, accommodating diverse data sources.

Initial Data Overview: Upon loading, the software automatically generates a statistical summary - including data dimensions, data types, missing value counts, and a preview table - providing an immediate snapshot of the dataset’s characteristics.

Advanced preprocessing with an intelligent assistant

Recognizing that data preprocessing is a critical and often complex step, a dedicated module with semi-automated, interactive capabilities has been developed.

Intelligent Data Quality Analyzer: This core component performs a multi-dimensional analysis of the dataset, evaluating completeness, uniqueness, validity, and consistency. It generates an overall data quality score and provides a prioritized list of actionable recommendations for remediation.

Interactive Cleaning Tools: The GUI provides tools for handling missing data and outliers. Users can select from a range of algorithms - from simple statistical methods (mean, median) to advanced techniques such as KNNImputer, IterativeImputer, and Isolation Forest - and preview their effects before application.

State Management and Reversibility: A built-in StateManager tracks every preprocessing operation. This provides full undo/redo functionality, empowering users to experiment with different cleaning strategies without the risk of irreversible changes.

Feature engineering and selection

This module offers a comprehensive suite of tools for feature engineering and selection, processes that are particularly crucial in materials science.

Multi-Strategy Feature Selection: A multi-stage feature selection workflow is implemented, allowing for a systematic reduction of dimensionality:
Importance-based Filtering: Utilizes model-intrinsic metrics (e.g., .feature_importances_, .coef_) for rapid, initial feature filtering.Correlation-based Filtering: Identifies and removes highly correlated features based on a user-defined threshold, while prioritizing the preservation of features that are more important for model performance.
Advanced Wrapper Methods: For more rigorous selection, users can employ advanced search algorithms, including genetic algorithms (GA) and recursive feature elimination (RFE), which evaluate feature subsets based on model performance.

Model training and hyperparameter optimization

The training module streamlines the process of constructing and validating predictive models.

Broad Model Library: It integrates a wide array of models from Scikit-learn^[12], eXtreme Gradient Boosting (XGBoost)^[13], Light Gradient Boosting Machine (LightGBM)^[14], and Categorical Boosting (CatBoost)^[15], supporting both regression and classification tasks. This diverse library was curated to include models, such as gradient boosting machines (XGBoost, LightGBM), known for their state-of-the-art performance on the structured, tabular data for which MatSci-ML Studio is designed.

Automated Hyperparameter Optimization: Hyperparameter tuning is automated using the Optuna library^[16], which employs efficient Bayesian optimization to identify optimal model configurations. Optuna was selected for its efficient Bayesian optimization and pruning algorithms. These are crucial for automating this otherwise complex and computationally expensive task, thereby making best-practice modeling accessible to non-experts.

Robust Evaluation: The module enforces best practices by performing all training and evaluation within a cross-validation (CV) framework to prevent data leakage and provide reliable performance estimates. It automatically generates key metrics and visualizations, such as confusion matrices, receiver operating characteristic (ROC) curves, and residual plots.

Model interpretability via SHapley Additive exPlanation analysis (shap_analysis.py)

To move beyond opaque “black-box” models, an integrated SHapley Additive exPlanations (SHAP) analysis module provides enhanced model interpretability.

Automated SHAP Calculation: Upon user request, the module calculates SHAP values to quantify the contribution of each feature to individual predictions. SHAP was selected due to its robust game-theoretic foundation and model-agnostic nature. This allows for a unified and reliable approach to interpreting any model within the library, which is a key requirement for building trust and generating scientific insight.

Interactive Visualizations: It automatically generates a suite of SHAP plots, including:Summary (Beeswarm) Plots: For a global view of feature importance and impact.Dependence Plots: To uncover complex, non-linear relationships between features and model output.Waterfall and Force Plots: For explaining individual predictions, enhancing local interpretability.

These visualizations are available both within the main interface and in separate, interactive windows.

Prediction, optimization, and active learning

MatSci-ML Studio provides a suite of advanced modules for the application of trained models.

Prediction: Users can load a trained model and apply it to new, unseen data, either by importing a file or via manual input for single-point predictions.Target Optimization: For inverse design problems, this module employs algorithms such as GA and Particle Swarm Optimization to find the optimal feature combinations that maximize or minimize a predicted target property.Multi-Objective Optimization: Addressing the common need to balance competing properties (e.g., strength vs. ductility), this module uses algorithms such as NSGA-II, Non-dominated Sorting Genetic Algorithm III (NSGA-III)^[17,18], Improved Strength Pareto Evolutionary Algorithm (SPEA2)^[19], Multi-objective Evolutionary Algorithm Based on Decomposition (MOEA/D)^[20], Derandomized Evolution Strategy with Covariance Matrix Adaptation (CMA-ES)^[21], Differential Evolution^[22], to identify the Pareto-optimal front, providing a set of non-dominated solutions. The inclusion of these powerful, well-established evolutionary algorithms directly addresses the complex, multi-faceted nature of real-world materials design problems. This enables users to move beyond single-property prediction toward practical, trade-off-aware optimization.Active Learning: To guide efficient experimental design, the active learning module uses Bayesian optimization to suggest the next most informative experiment to perform. It intelligently balances exploration (reducing model uncertainty) with exploitation (improving the target property).

Performance monitoring and resource management

To ensure a stable and efficient user experience, the software incorporates a real-time performance monitor. This module tracks system resources, such as central processing unit (CPU) and memory utilization, and provides real-time feedback on task progress. It also issues alerts to prevent system overload, which is particularly critical during computationally intensive tasks such as hyperparameter optimization or feature selection.

Case 1: strength prediction of aluminum alloys

In alloy property prediction, raw compositional data are seldom used directly as input features because they lack sufficient physical and chemical information^[4,23]. Elemental percentages alone cannot capture critical factors such as thermodynamic interactions, atomic size mismatches, and electronic structure variations among constituent elements. To address this limitation, four feature mapping strategies were adopted to construct corresponding alloy descriptors, as detailed in previous studies^[24-26]:

where a_i represents the atomic fraction of the i-th element, and x_i denotes a specific elemental property (e.g., atomic radius or electronegativity). N is the total number of elements in the alloy. Four types of descriptors are derived: the weighted average (x₁), the harmonic mean (x₂), the weighted variance (x_d), which reflects property dispersion, and the range of normalized squared deviations (x_r), which quantifies the degree of property mismatch among elements. In total, 66 elemental properties were considered, resulting in 264 input features and one target variable (UTS). These descriptors are widely adopted in feature engineering for materials informatics.

The automated workflow within MatSci-ML Studio and its comprehensive outputs are illustrated in Figure 2. To manage the high-dimensional input space of 264 features, the process begins with an automated feature importance analysis. Specifically, a cross-validated XGBoost model is employed to rank all features by their predictive power, with the results visualized in Figure 2A. This initial step immediately provides valuable domain-specific insight, revealing that descriptors related to ionic radii [e.g., x2_S2_radii_ionic_(Yagoda)_(A)] are the most influential factors in determining UTS. The software then addresses the common issue of multicollinearity, visualized in the initial correlation heatmap [Figure 2B], where dense clusters of highly correlated features are evident. By applying a performance-aware correlation filtering algorithm (Pearson’s r > 0.95), the system prunes redundant features. This reduces the feature set to a robust subset of 13 descriptors, a result confirmed by the significantly sparser correlation matrix in Figure 2C.

Figure 2. Comprehensive visualization of the automated regression workflow in MatSci-ML Studio for predicting aluminum alloy strength. (A) Feature importance plot ranking the most influential descriptors based on their predictive power. Error bars indicate the standard deviation of importance scores across CV folds; (B) Initial correlation matrix revealing multicollinearity among the 264 engineered features; (C) Correlation matrix of the refined subset of 13 features after automated filtering, displaying significantly reduced inter-correlation; (D) Progress of Bayesian hyperparameter optimization over 30 iterations, plotting the CV R² score and indicating the best score achieved (red dashed line); (E) Distribution of CV scores from the optimization search, confirming the stability of the optimal parameter set; (F) Parameter vs. score plot for the colsample_bytree hyperparameter, illustrating the optimization search path; (G) Residuals plot for the final model, showing the difference between predicted and actual values for both training (blue) and test (red) sets; (H) Distribution of residuals, centered around zero, indicating an unbiased model; (I) Predicted vs. actual UTS values for the training (R² = 0.987) and test (R² = 0.915) sets. The dashed black line represents perfect prediction (y = x). ML: Machine learning; CV: cross-validation.

With a refined feature set, the workflow seamlessly transitions to hyperparameter optimization. MatSci-ML Studio utilizes a Bayesian search strategy, and its progress over 30 iterations is illustrated in Figure 2D. The plot reveals a clear convergence toward an optimal hyperparameter configuration, where the best CV R² score reached approximately 0.8468. The score distribution across all trials [Figure 2E] and the parameter-versus-score plot [Figure 2F] further validate the stability and effectiveness of the optimization process, confirming that the system successfully navigated the hyperparameter landscape to identify a high-performing model configuration.

The final stage involves training the optimized XGBoost model on the selected features and evaluating its performance on a held-out test set. The predictive accuracy is powerfully illustrated in the predicted-versus-actual plot [Figure 2I], where the model achieves an outstanding R² of 0.915 and a root mean square error (RMSE) of 47.7 MPa on the test set. This performance closely mirrors the training set metrics (R² = 0.987, RMSE = 26.1 MPa), indicating strong generalization capabilities. Furthermore, the residuals plot and its corresponding distribution [Figure 2G and H] exhibit a random pattern centered around zero, confirming that the model has captured the underlying data trends without systematic bias. These results are consistent with those reported by Jiang et al., who achieved an R² of 0.93 on the same test dataset using a comparable methodology^[25]. This validates that the MatSci-ML Studio automated pipeline can construct high-performing regression models that are competitive with manually curated workflows.

Beyond predictive accuracy, MatSci-ML Studio provides deeper scientific insight via an automated SHAP analysis [Figure 3]. The global feature importance plot [Figure 3A], which ranks features by their mean absolute SHAP value, offers a model-agnostic confirmation of the primary physical drivers of UTS. This validates that descriptors related to atomic and ionic properties - notably ionic radii [x2_S2_radii_ionic_(Yagoda)_(A)] and lattice constant [x1_S11_lattice_constant_a_(pm)] - are the most influential factors. Moving from global importance to local explanations, the SHAP summary plot [Figure 3B] visualizes how each feature’s value impacts individual predictions. For instance, it reveals a clear positive correlation for x2_S2_radii_ionic_(Yagoda)_(A), where higher feature values (red points) consistently yield positive SHAP values, indicating a contribution to increased strength. Conversely, a feature such as x_r_E5_energy_ionization_first exhibits an inverse trend, where lower values (blue points) positively impact the predicted strength. Furthermore, the plot uncovers complex interaction effects; the wide vertical dispersion of SHAP values for x1_S11_lattice_constant_a_(pm) suggests that its influence is highly context-dependent and modulated by other features. This dual-level interpretability, generated automatically, empowers researchers not only to trust model predictions but also to formulate new hypotheses about the underlying material physics.

Figure 3. SHAP analysis results for interpreting the final XGBoost model predicting aluminum alloy strength. (A) Global feature importance ranked by mean absolute SHAP value, indicating the average magnitude of impact on model predictions; (B) SHAP summary plot (beeswarm) illustrating the distribution, direction (positive/negative SHAP value), and magnitude of impact for each feature across individual predictions, colored by the feature’s original value (red = high, blue = low). SHAP: SHapley Additive exPlanations; XGBoost: eXtreme Gradient Boosting.

Case 2: lung cancer prediction

To demonstrate the framework’s versatility beyond regression and its capability to handle heterogeneous data, MatSci-ML Studio was applied to a publicly available lung cancer prediction dataset. This dataset contains a mix of categorical (e.g., “SMOKING”) and numerical features, making it an excellent test case for our automated preprocessing and classification pipeline. The complete set of results is summarized in Figure 4.

Figure 4. Complete workflow visualization generated by MatSci-ML Studio for the lung cancer prediction (classification) task. (A) Feature importance plot, where SMOKING is identified as the most influential factor. The importance of one-hot encoded features has been aggregated back to the original categorical feature; (B) Initial correlation matrix showing low inter-correlation among features; (C) Correlation matrix after filtering, demonstrating that most original features were retained; (D) Progress of Bayesian hyperparameter optimization maximizing the F1-score, which converges to a best value of approximately 0.9123; (E) Score distribution from the hyperparameter search; (F) Relationship between the min_samples_leaf hyperparameter and the CV score; (G) ROC curve for the final model on the test set, achieving an AUC of 0.917; (H) Precision-Recall curve, showing a high average precision (AP = 0.885); (I) Confusion matrix for the test set, detailing the classification performance (TN = 544, FP = 49, FN = 51, TP = 356). ML: Machine learning; CV: cross-validation; ROC: receiver operating characteristic; AUC: area under the curve.

The workflow, similar to Case 1, commences with feature analysis. The feature importance plot [Figure 4A], which aggregates the impact of one-hot encoded variables, immediately identifies SMOKING as the most critical predictive factor. The initial correlation heatmap [Figure 4B] reveals very low multicollinearity among the features, suggesting that most provide independent information. Consequently, the automated correlation filtering step [Figure 4C] retains the majority of the features, a decision intelligently made by the system based on the data’s statistical properties.

Even with relatively independent features, hyperparameter optimization remains crucial for maximizing model performance. MatSci-ML Studio’s automated Bayesian search process, visualized in Figure 4D, effectively maximizes the F1 score, converging to a peak CV score of approximately 0.9123. The stability of this optimum is confirmed by the score distribution plot [Figure 4E], while Figure 4F offers a granular view of the search, plotting the CV score against the min_samples_leaf hyperparameter. The final evaluation of the optimized classifier on unseen test data showcases its excellent performance. The ROC curve [Figure 4G] yields an area under the curve (AUC) of 0.917, indicating excellent class separability. This is corroborated by the Precision-Recall curve [Figure 4H], which demonstrates a high average precision of 0.885, particularly important for imbalanced datasets. The confusion matrix [Figure 4I] details the model’s performance, recording 544 true negatives and 356 true positives against a relatively small number of misclassifications.

In summary, the automated workflow achieved an accuracy of 90% and an F1-score of 88%, which represents excellent predictive performance on this dataset. These results closely align with those reported by Ahmed [Figure 5], further validating the applicability and reliability of the proposed framework^[27].

Figure 5. Confusion matrix of the classification model on lung cancer prediction dataset obtained by Ahmed^[27].

To ensure the final classifier is not merely a “black box”, MatSci-ML Studio integrates an automated SHAP analysis to provide transparent and interpretable insights into the model’s decision-making process. The results, shown in Figure 6, illuminate the key factors driving the lung cancer predictions. The global feature importance plot [Figure 6A], ranked by the mean absolute SHAP value, unequivocally identifies smoking-related features as the most dominant predictors. The one-hot encoded features SMOKING_0 (non-smoker) and SMOKING_1 (smoker) exert the largest average impact on the model’s output, far surpassing other factors such as ENERGY_LEVEL and THROAT_DISCOMFORT_0. This provides immediate, quantifiable evidence of the most critical risk factors learned by the model from the data.

Figure 6. SHAP interpretability analysis for the final classification model. (A) Global feature importance ranked by mean absolute SHAP value (|SHAP value|), indicating the average impact magnitude on the model output; (B) SHAP summary plot (beeswarm) showing the distribution of SHAP values for each feature across individual predictions. Each point represents a single prediction; its position on the x-axis denotes the impact on the model output (positive values push toward class 1, negative toward class 0), and its color represents the original feature value (red = high/present, blue = low/absent). SHAP: SHapley Additive exPlanations.

However, the SHAP summary plot [Figure 6B] provides a richer, more directional understanding. Each point on this beeswarm plot corresponds to a single prediction, colored by the feature’s value (red for present/high, blue for absent/low). This visualization clearly reveals the learned relationships. For instance, the presence of SMOKING_1 (value = 1, red points) consistently produces large positive SHAP values, strongly pushing the prediction toward the “Cancer” class (class 1). Conversely, the presence of SMOKING_0 (red points, which means the patient is a non-smoker) yields strongly negative SHAP values, driving the prediction towards the “No Cancer” class (class 0).

This plot also uncovers more subtle patterns. For ENERGY_LEVEL, a clear inverse relationship is visible: high energy levels (red points) are associated with negative SHAP values (lower cancer risk), while low energy levels (blue points) contribute positively to the prediction (higher cancer risk). Similarly, the presence of THROAT_DISCOMFORT_1 and BREATHING_ISSUE_1 (not shown in the top features but visible in the beeswarm) consistently pushes predictions towards a positive diagnosis. The vertical spread of SHAP values for a given feature, such as EXPOSURE_TO_POLLUTION_1, indicates interaction effects, where the impact of pollution exposure may be amplified or mitigated by other factors. By automatically generating and presenting this dual-level interpretability, MatSci-ML Studio enables users to validate that the model has learned medically intuitive and scientifically sound relationships, thereby fostering trust in its predictions.

Synthesis and implications from case studies

The two case studies, while demonstrating the versatility of MatSci-ML Studio, originate from fundamentally different contexts, as astutely noted by our reviewers. Case 1 (Alloy UTS) exemplifies a typical materials science problem that involves what can be termed the “compression of a hierarchical structure”, where low-level physical descriptors are engineered to predict a high-level macroscopic property. In contrast, Case 2 (Lung Cancer) represents a clinical problem where features exist on a “single hierarchical level” of empirical observation.

This distinction highlights a key strength and a clear limitation of our toolkit. The utility of MatSci-ML Studio lies in its domain-agnostic architecture: it provides a robust, automated workflow for any problem that can be formulated with structured, tabular data, regardless of feature origin. Its limitation, however, is that it operates on this final tabular representation and does not automate the complex, domain-specific process of feature engineering itself (e.g., deriving the physical descriptors in Case 1). The toolkit empowers researchers once these foundational features are established, accelerating the subsequent steps of model building, interpretation, and optimization.

Advantages

It is crucial to first clarify that MatSci-ML Studio, as an induction-based toolkit, operates on correlations within data and does not inherently model underlying physical or biological causal mechanisms. Its primary role is to serve as a convenient and powerful platform for researchers to strategically execute data-driven discovery. Within this well-defined scope, MatSci-ML Studio represents a significant methodological advancement.

MatSci-ML Studio delivers significant methodological advantages by targeting three key challenges in computational materials science: accessibility, reproducibility, and analytical depth. First, it champions accessibility. Unlike traditional script-based frameworks, MatSci-ML Studio is built entirely around an intuitive GUI. This no-code design fundamentally democratizes advanced ML. It empowers domain experts to perform sophisticated analyses without the hurdle of extensive programming, thereby shifting their focus from technical implementation to scientific inquiry. Second, it enhances end-to-end reproducibility. We address the common challenge of fragmented workflows by encapsulating the entire research pipeline - from data ingestion to model interpretation - within a single, cohesive application. This seamless integration is powerfully enhanced by a built-in Project Management and Version Control system, a feature that facilitates traceable and reproducible research by allowing users to snapshot and restore entire experimental states.

Furthermore, the software is designed with embedded intelligence to guide the user. The Advanced Preprocessing module, for instance, does not merely provide a set of tools; it includes an EnhancedDataQualityAnalyzer that automatically assesses the dataset and offers actionable, prioritized recommendations. This semi-automated guidance, combined with an interactive interface featuring an undo/redo state manager, fosters an exploratory and risk-free environment for data preparation. This approach contrasts sharply with the “black-box” nature of fully automated scripts, placing the domain expert firmly in control while alleviating the technical burden.

Finally, the most significant strategic contribution of MatSci-ML Studio is its ability to logically integrate multiple tools into a seamless, end-to-end design loop. The platform transforms the ML model from a passive predictor into a proactive engine for discovery. This is achieved by integrating the forward modeling workflow with inverse design capabilities. Researchers can first build and validate a predictive model, then use the integrated SHAP analysis to understand its driving factors, and finally, employ this same model as a high-fidelity surrogate within the target optimization and multi-objective optimization modules. This seamless chaining of analysis, interpretation, and optimization is the core of the toolkit’s strategic power, enabling researchers to efficiently navigate complex design spaces and accelerate the discovery of novel materials.