Interpretable Machine Learning Research Articles

Diffusion model approach to simulating electron-proton scattering events

Generative artificial intelligence is a fast-growing area of research offering various avenues for exploration in high-energy nuclear physics. In this work, we explore the use of generative models for simulating electron-proton collisions relevant to experiments like the Continuous Electron Beam Accelerator Facility and the future Electron-Ion Collider (EIC). These experiments play a critical role in advancing our understanding of nucleons and nuclei in terms of quark and gluon degrees of freedom. The use of generative models for simulating collider events faces several challenges such as the sparsity of the data, the presence of global or eventwide constraints, and steeply falling particle distributions. In this work, we focus on the implementation of diffusion models for the simulation of electron-proton scattering events at EIC energies. Our results demonstrate that diffusion models can reproduce relevant observables such as momentum distributions and correlations of particles, momentum sum rules, and the leading electron kinematics, all of which are of particular interest in electron-proton collisions. Although the sampling process is relatively slow compared to other machine-learning architectures, we find diffusion models can generate high-quality samples. We foresee various applications of our work including inference for nuclear structure, interpretable generative machine learning, and searches of physics beyond the Standard Model. Published by the American Physical Society 2024

Interpretable machine learning models for predicting the bond strength between UHPC and normal-strength concrete

Ultra-high-performance concrete (UHPC) boasts superior mechanical properties and durability, rendering it a viable material to repair and reinforce pre-existing concrete structures. The complexity of the bond behavior between UHPC and normal-strength concrete (NC) presents a challenge to make precise predictions of the interfacial bond strength. This study employed four machine learning algorithms, namely artificial neural network, random forest, adaptive boosting, and categorical gradient boosting, to predict the interfacial bond strength of UHPC-NC. A database of 95 samples of slant shear tests was collected from existing literature. The input variables for modeling include joint angle, compressive strength of NC and UHPC, surface treatment, interfacial moisture, curing age, and curing method. The impact of each feature on the slant shear bond strength was assessed using SHapley Additive exPlanations and Partial Dependence Plots. Results show that the categorical gradient boosting model performed the best on the testing set, achieving an R2 of 0.948, RMSE of 1.408, and MAE of 1.028. The accuracy of all four machine learning models surpassed that of two empirical models (AASHTO LRFD 2014 and Eurocode 2). The top three features influencing the interfacial bond strength were identified as the surface treatment, joint angle, and compressive strength of NC, while the impact of the curing method and the interfacial moisture were comparatively limited. The proposed machine learning model exhibits potential as a valuable tool for assisting structural retrofitting and rehabilitation with UHPC.

Understanding stacking fault energy of NbMoTaW multi-principal element alloys by interpretable machine learning

Stacking fault energy (SFE) is a crucial property influencing the deformation mechanisms of multi-principal element alloys (MPEAs). However, experimentally measuring SFE and exploring composition space of MPEAs is challenging. This study explores the SFE in NbMoTaW by integrating machine learning force fields (ML-FF), molecular dynamics simulations (MD), neural networks, and symbolic regression (SR) methods. A SFE dataset containing 2000 different NbMoTaW compositions were generated by combining ML-FF and MD. Then the neural-network model was developed for SFE prediction with an accuracy of 0.981 and a mean absolute error of 0.020. Using SR, the valence electron concentration (VEC), average shear modulus (G), radii gamma (γ) were identified as the key descriptors to determine SFE with the relationship of SFE=0.307·VEC+0.352·(G+γ). This work demonstrated a significant impact of chemical composition on SFE and established an accurate mathematical expression for SFE prediction, enhancing the understanding and alloy design of MPEAs.

Prediction and analysis of mechanical properties of hot-rolled strip steel based on an interpretable machine learning

Establishing a low-cost, high-precision predictive model for mechanical properties is crucial for enhancing the properties of steel. Due to the complexity of the steel production process, traditional mechanistic models struggle to efficiently and accurately describe the relationships among alloy elements, processes, and properties. In response to this, the paper introduces a high-precision framework for predicting and optimizing mechanical properties through feature engineering and machine learning models. Firstly, an effective dimension reduction of features is achieved through a weighted fusion heterogeneous feature selection strategy, thereby identifying the optimal model input features. Subsequently, an improved seagull optimization algorithm is utilized to optimize the hyperparameters of Extreme Gradient Boosting, further enhancing the predictive accuracy of the model. Moreover, based on the Shapley Additive Explanation method, a quantitative analysis of the model's predictive outcomes is conducted, elucidating the impacts of alloy elements and processes on mechanical properties, consistent with established principles of physical metallurgy. The proposed framework not only enables precise prediction of mechanical properties in steel but also provides theoretical guidance and technical support for process optimization design and the development of new steel grades.

Interpretable machine learning models for predicting clinical pregnancies associated with surgical sperm retrieval from testes of different etiologies: a retrospective study

BackgroundThe relationship between surgical sperm retrieval of different etiologies and clinical pregnancy is unclear. We aimed to develop a robust and interpretable machine learning (ML) model for predicting clinical pregnancy using the SHapley Additive exPlanation (SHAP) association of surgical sperm retrieval from testes of different etiologies.MethodsA total of 345 infertile couples who underwent intracytoplasmic sperm injection (ICSI) treatment with surgical sperm retrieval due to different etiologies from February 2020 to March 2023 at the reproductive center were retrospectively analyzed. The six machine learning (ML) models were used to predict the clinical pregnancy of ICSI. After evaluating the performance characteristics of the six ML models, the Extreme Gradient Boosting model (XGBoost) was selected as the best model, and SHAP was utilized to interpret the XGBoost model for predicting clinical pregnancies and to reveal the decision-making process of the model.ResultsCombining the area under the receiver operating characteristic curve (AUROC), accuracy, precision, recall, F1 score, brier score, and the area under the precision-recall (P-R) curve (AP), the XGBoost model has the best performance (AUROC: 0.858, 95% confidence interval (CI): 0.778–0.936, accuracy: 79.71%, brier score: 0.151). The global summary plot of SHAP values shows that the female age is the most important feature influencing the model output. The SHAP plot showed that younger age in females, bigger testicular volume (TV), non-tobacco use, higher anti-müllerian hormone (AMH), lower follicle-stimulating hormone (FSH) in females, lower FSH in males, the temporary ejaculatory disorders (TED) group, and not the non-obstructive azoospermia (NOA) group all resulted in an increased probability of clinical pregnancy.ConclusionsThe XGBoost model predicts clinical pregnancies associated with testicular sperm retrieval of different etiologies with high accuracy, reliability, and robustness. It can provide clinical counseling decisions for patients with surgical sperm retrieval of various etiologies.

Interpretable machine learning boosting the discovery of targeted organometallic compounds with optimal bandgap

Organometallic compounds (OMCs) have attracted tremendous attention in various fields, such as photovoltaic cell and high-k dielectric application, due to their beneficial properties. Despite their potential, the progression of OMCs into industrial applications is hindered by the limited databases available for their properties and the absence of efficient surrogate models. To address this, in this study, optimally selected feature-based surrogate models for predicting the electronic properties of OMCs are constructed via various multiscale features and extensive database. To this end, high-throughput calculation was performed to obtain electronic properties of more than 18k materials generally known as organometallics, augmenting around 12k organic materials obtained from the public open data set, OMDB-GAP1. For generating features closely related to OMCs, descriptors encapsulating the information ranging local to global, also other widely-used composition-, structure-based features (more than 3.5k in total) were employed. Among these descriptors, we identified 48 critical features that elucidates the physicochemical underpinnings of OMCs’ properties, suggesting their impact on the properties of OMCs. The light gradient boosting machine model achieved high-accuracy predictions across the entire database with just 1 % of the total descriptors, sufficiently compared to the entire sets (decreased of around 0.01 by R2 score and 0.01 eV by MAE). Furthermore, the efficacy of active learning process was demonstrated to find OMCs with optimal properties rapidly. As a result, expected improvement outperforms other methods by identifying 69 % of the target materials only searching 46 % of the total search space. Our constructed platform with a high-throughput calculated database can pave the way for the rapid screening of OMCs for the targeted industrial application, and suggest a comprehensive grasp of the intrinsic properties of OMCs and related compounds.

Predicting graft and patient outcomes following kidney transplantation using interpretable machine learning models

The decision to accept a deceased donor organ offer for transplant, or wait for something potentially better in the future, can be challenging. Clinical decision support tools predicting transplant outcomes are lacking. This project uses interpretable methods to predict both graft failure and patient death using data from previously accepted kidney transplant offers. Using more than 25 years of transplant outcome data, we train and compare several survival analysis models in single risk settings. In addition, we use post hoc interpretability techniques to clinically validate these models. Neural networks show comparable performance to the Cox proportional hazard model, with concordance of 0.63 and 0.79 for prediction of graft failure and patient death, respectively. Donor and recipient ages, the number of mismatches at DR locus, dialysis type, and primary renal disease appear to be important features for transplant outcome prediction. Owing to their good predictive performance and the clinical relevance of their post hoc interpretation, neural networks represent a promising core component in the construction of future decision support systems for transplant offering.

Interpretable Machine Learning for Predicting Heart Failure Following Acute Coronary Syndrome in Patients With Metabolic Dysfunction-Associated Steatotic Liver Disease

Interpretable Machine Learning for Predicting Heart Failure Following Acute Coronary Syndrome in Patients With Metabolic Dysfunction-Associated Steatotic Liver Disease

Interpretable machine learning guided by physical mechanisms reveals drivers of runoff under dynamic land use changes

Human activities continuously impact water balances and cycling in watersheds, making it essential to accurately identify the responses of runoff to dynamic changes in land use types. Although machine learning models demonstrate promise in capturing the intricate interplay between hydrological factors, their “black box” nature makes it challenging to identify the dynamic drivers of runoff. To overcome this challenge, we employed an interpretable machine learning method to inversely deduce the dynamic determinants within hydrological processes. In this study, we analyzed land use changes in the Ningxia section of the middle Yellow River across four periods, laying the foundation for revealing how these changes affect runoff. The sub-watershed attributes and meteorological characteristics generated by the Soil and Water Assessment Tool (SWAT) model were used as input variables of the Extreme Gradient Boosting (XGBoost) model to simulate substantial sub-watershed rainfall runoff in the region. The XGBoost was interpreted using the SHapley Additive exPlanations (SHAP) to identify the dynamic responses of runoff to the land use changes over different periods. The results revealed increasingly frequent interchanges between the land use types in the study area. The XGBoost effectively captured the characteristics of the hydrological processes in the SWAT-derived sub-watersheds. The SHAP analysis results demonstrated that the promoting effect of agricultural land (AGRL) on runoff gradually weakens, while forests (FRST) continuously strengthen their restraining effect on runoff. Relevant land use policies provide empirical support for these findings. Furthermore, the interaction between meteorological variables and land use impacts the runoff generation mechanism and exhibits a threshold effect, with the thresholds for relative humidity (RH), maximum temperature (MaxT), and minimum temperature (MinT) determined to be 0.8, 25 °C, and 15 °C, respectively. This reverse deduction method can reveal hydrological patterns and the mechanisms of interaction between variables, helping to effectively addressing constantly changing human activities and meteorological conditions.

Development and Validation of an Interpretable Machine Learning Model for Early Prognosis Prediction in ICU Patients with Malignant Tumors and Hyperkalemia.

This study aims to develop and validate a machine learning (ML) predictive model for assessing mortality in patients with malignant tumors and hyperkalemia (MTH). We extracted data on patients with MTH from the Medical Information Mart for Intensive Care-IV, version 2.2 (MIMIC-IV v2.2) database. The dataset was split into a training set (75%) and a validation set (25%). We used the Least Absolute Shrinkage and Selection Operator (LASSO) regression to identify potential predictors, which included clinical laboratory indicators and vital signs. Pearson correlation analysis tested the correlation between predictors. In-hospital death was the prediction target. The Area Under the Curve (AUC) and accuracy of the training and validation sets of 7 ML algorithms were compared, and the optimal 1 was selected to develop the model. The calibration curve was used to evaluate the prediction accuracy of the model further. SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) enhanced model interpretability. 496 patients with MTH in the Intensive Care Unit (ICU) were included. After screening, 17 clinical features were included in the construction of the ML model, and the Pearson correlation coefficient was <0.8, indicating that the correlation between the clinical features was small. eXtreme Gradient Boosting (XGBoost) outperformed other algorithms, achieving perfect scores in the training set (accuracy: 1.000, AUC: 1.000) and high scores in the validation set (accuracy: 0.734, AUC: 0.733). The calibration curves indicated good predictive calibration of the model. SHAP analysis identified the top 8 predictive factors: urine output, mean heart rate, maximum urea nitrogen, minimum oxygen saturation, minimum mean blood pressure, maximum total bilirubin, mean respiratory rate, and minimum pH. In addition, SHAP and LIME performed in-depth individual case analyses. This study demonstrates the effectiveness of ML methods in predicting mortality risk in ICU patients with MTH. It highlights the importance of predictors like urine output and mean heart rate. SHAP and LIME significantly enhanced the model's interpretability.

Understanding the mechanism of gully erosion in the alpine region through an interpretable machine learning approach

In the alpine region, climate warming has led to the retreat of glaciers, snow cover, and permafrost. This has intensified water cycling, soil erosion, and increased the occurrence of natural disasters in the alpine region. This study investigated the Lhasa River Basin in the southern Tibetan Plateau, serving as a representative case study of a typical alpine basin, with a specific focus on gully erosion. Based on field investigations and interpretation using high-resolution satellite remote sensing images, the Random Forest (RF) algorithm was applied to evaluate gully erosion susceptibility on watershed level. The Shapley Additive Interpretation method was then used to interpret the RF model and gain deeper insights into the influencing variables of gully erosion. The results showed that the RF model achieved an area under the receiver operating characteristic (AUC) accuracy of 0.99 and 0.98 for the training and testing datasets, respectively, indicating an outstanding performance of the model. The resulting susceptibility map based on the RF model shows that areas with moderate and higher levels of gully erosion susceptibility are covering 50 % of the basin. The model interpretation results indicated that elevation, slope, permafrost, rainstorm, silt loam topsoil, human activity, stream power, and vegetation were the explaining variables with the highest importance for gully erosion occurrence. Different variables are characterized by specific thresholds promoting gully erosion such as: i) elevations higher than 4950 m, ii) slopes steeper than 13.5°, iii) extreme rainstorms longer than 11 days per year, iv) silt loam topsoil, v) presence of permafrost, vi) stream power index higher than 1.2, and vii) normalized difference vegetation index (NDVI) lower than 0.25. Our findings provide the scientific basis to improve soil erosion control in such highly vulnerable alpine area.

Machine learning-enabled prediction of prolonged length of stay in hospital after surgery for tuberculosis spondylitis patients with unbalanced data: a novel approach using explainable artificial intelligence (XAI)

BackgroundTuberculosis spondylitis (TS), commonly known as Pott’s disease, is a severe type of skeletal tuberculosis that typically requires surgical treatment. However, this treatment option has led to an increase in healthcare costs due to prolonged hospital stays (PLOS). Therefore, identifying risk factors associated with extended PLOS is necessary. In this research, we intended to develop an interpretable machine learning model that could predict extended PLOS, which can provide valuable insights for treatments and a web-based application was implemented.MethodsWe obtained patient data from the spine surgery department at our hospital. Extended postoperative length of stay (PLOS) refers to a hospitalization duration equal to or exceeding the 75th percentile following spine surgery. To identify relevant variables, we employed several approaches, such as the least absolute shrinkage and selection operator (LASSO), recursive feature elimination (RFE) based on support vector machine classification (SVC), correlation analysis, and permutation importance value. Several models using implemented and some of them are ensembled using soft voting techniques. Models were constructed using grid search with nested cross-validation. The performance of each algorithm was assessed through various metrics, including the AUC value (area under the curve of receiver operating characteristics) and the Brier Score. Model interpretation involved utilizing methods such as Shapley additive explanations (SHAP), the Gini Impurity Index, permutation importance, and local interpretable model-agnostic explanations (LIME). Furthermore, to facilitate the practical application of the model, a web-based interface was developed and deployed.ResultsThe study included a cohort of 580 patients and 11 features include (CRP, transfusions, infusion volume, blood loss, X-ray bone bridge, X-ray osteophyte, CT-vertebral destruction, CT-paravertebral abscess, MRI-paravertebral abscess, MRI-epidural abscess, postoperative drainage) were selected. Most of the classifiers showed better performance, where the XGBoost model has a higher AUC value (0.86) and lower Brier Score (0.126). The XGBoost model was chosen as the optimal model. The results obtained from the calibration and decision curve analysis (DCA) plots demonstrate that XGBoost has achieved promising performance. After conducting tenfold cross-validation, the XGBoost model demonstrated a mean AUC of 0.85 ± 0.09. SHAP and LIME were used to display the variables’ contributions to the predicted value. The stacked bar plots indicated that infusion volume was the primary contributor, as determined by Gini, permutation importance (PFI), and the LIME algorithm.ConclusionsOur methods not only effectively predicted extended PLOS but also identified risk factors that can be utilized for future treatments. The XGBoost model developed in this study is easily accessible through the deployed web application and can aid in clinical research.

Prediction of Compressive Strength of Concrete Specimens Based on Interpretable Machine Learning.

The aim of this paper is to explore an effective model for predicting the compressive strength of concrete using machine learning technology, as well as to interpret the model using an interpretable method, which overcomes the limitation of the unknowable prediction processes of previous machine learning models. An experimental database containing 228 samples of the compressive strength of standard cubic specimens was built in this study, and six algorithms were applied to build the predictive model. The results show that the XGBoost model has the highest prediction accuracy among all models, as the R2 of the training set and testing set are 0.982 and 0.966, respectively. Further analysis was conducted on the XGBoost model to discuss its applicability. The main steps include the following: (i) obtaining key features, (ii) obtaining trends in the evolution of features, (iii) single-sample analysis, and (iv) conducting a correlation analysis to explore methods of visualizing the variations in the factors that exert influence. The interpretability analyses on the XGBoost model show that the contribution to the compressive strength by each factor is highly in line with the conventional theory. In summary, the XGBoost model proved to be effective in predicting concrete's compressive strength.

Predicting Intimate Partner Violence Perpetration Among Young Adults Experiencing Homelessness in Seven U.S. Cities Using Interpretable Machine Learning.

Young adults experiencing homelessness (YAEH) are at higher risk for intimate partner violence (IPV) victimization than their housed peers. This is often due to their increased vulnerability to abuse and victimization before and during homelessness, which can result in a cycle of violence in which YAEH also perpetrates IPV. Identifying and addressing factors contributing to IPV perpetration at an early stage can reduce the risk of IPV. Yet to date, research examining YAEH's IPV perpetration is scarce and has largely employed conventional statistical approaches that are limited in modeling this complex phenomenon. To address these gaps, this study used an interpretable machine learning approach to answer the research question: What are the most salient predictors of IPV perpetration among a large sample of YAEH in seven U.S. cities? Participants (N = 1,426) on average were 21 years old (SD = 2.09) and were largely cisgender males (59%) and racially/ethnically diverse (81% were from historically excluded racial/ethnic groups; i.e., African American, Latino/a, American Indian, Asian or Pacific Islander, and mixed race/ethnicity). Over one-quarter (26%) reported IPV victimization, and 20% reported IPV perpetration while homeless. Experiencing IPV victimization while homeless was the most important factor in predicting IPV perpetration. An additional 11 predictors (e.g., faced frequent discrimination) were positively associated with IPV perpetration, whereas 8 predictors (e.g., reported higher scores of mindfulness) were negatively associated. These findings underscore the importance of developing and implementing effective interventions with YAEH that can prevent IPV, particularly those that recognize the positive association between victimization and perpetration experiences.

Interpretable machine learning model based on clinical factors for predicting muscle radiodensity loss after treatment in ovarian cancer.

Muscle radiodensity loss after surgery and adjuvant chemotherapy is associated with poor outcomes in ovarian cancer. Assessing muscle radiodensity is a real-world clinical challenge owing to the requirement for computed tomography (CT) with consistent protocols and labor-intensive processes. This study aimed to use interpretable machine learning (ML) to predict muscle radiodensity loss. This study included 723 patients with ovarian cancer who underwent primary debulking surgery and platinum-based chemotherapy between 2010 and 2019 at two tertiary centers (579 in cohort 1 and 144 in cohort 2). Muscle radiodensity was assessed from pre- and post-treatment CT acquired with consistent protocols, and a decrease in radiodensity ≥ 5% was defined as loss. Six ML models were trained, and their performances were evaluated using the area under the curve (AUC) and F1-score. The SHapley Additive exPlanations (SHAP) method was applied to interpret the ML models. The CatBoost model achieved the highest AUC of 0.871 (95% confidence interval, 0.870-0.874) and F1-score of 0.688 (95% confidence interval, 0.685-0.691) among the models in the training set and outperformed in the external validation set, with an AUC of 0.839 and F1-score of 0.673. Albumin change, ascites, and residual disease were the most important features associated with a higher likelihood of muscle radiodensity loss. The SHAP force plot provided an individualized interpretation of model predictions. An interpretable ML model can assist clinicians in identifying ovarian cancer patients at risk of muscle radiodensity loss after treatment and understanding the contributors of muscle radiodensity loss.

Inferring dynamical models from time-series biological data using an interpretable machine learning method based on weighted expression trees

The growing time-series data make it possible to glimpse the hidden dynamics in various fields. However, developing a computational toolbox with high interpretability to unveil the interaction dynamics from data remains a crucial challenge. Here, we propose a new computational approach called automated dynamical model inference based on expression trees (ADMIET), in which the machine learning algorithm, the numerical integration of ordinary differential equations and the interpretability from prior knowledge are embedded into the symbolic learning scheme to establish a general framework for revealing the hidden dynamics in time-series data. ADMIET takes full advantage of both machine learning algorithm and expression tree. Firstly, we translate the prior knowledge into constraints on the structure of expression tree, reducing the search space and enhancing the interpretability. Secondly, we utilize the proposed adaptive penalty function to ensure the convergence of gradient descent algorithm and the selection of the symbols. Compared to gene expression programming, ADMIET exhibits its remarkable capability in function fitting with higher accuracy and broader applicability. Moreover, ADMIET can better fit parameters in nonlinear forms compared to regression methods. Furthermore, we apply ADMIET to two typical biological systems and one real data with different prior knowledge to infer the dynamical equations. The results indicate that ADMIET can not only discover the interaction relationships but also provide accurate estimates of the parameters in the equations. These results demonstrate ADMIET’s superiority in revealing interpretable dynamics from time-series biological data.

Construction and validation of predictive models for intravenous immunoglobulin-resistant Kawasaki disease using an interpretable machine learning approach.

Intravenous immunoglobulin (IVIG)-resistant Kawasaki disease is associated with coronary artery lesion development. This study aimed to explore the factors associated with IVIG-resistance and construct and validate an interpretable machine learning (ML) prediction model in clinical practice. Between December 2014 and November 2022, 602 patients were screened and risk factors for IVIG-resistance investigated. Five ML models are used to establish an optimal prediction model. The SHapley Additive exPlanations (SHAP) method was used to interpret the ML model. Na+, hemoglobin (Hb), C-reactive protein (CRP), and globulin were independent risk factors for IVIG-resistance. A nonlinear relationship was identified between globulin level and IVIG-resistance. The XGBoost model exhibited excellent performance, with an area under the receiver operating characteristic curve of 0.821, accuracy of 0.748, sensitivity of 0.889, and specificity of 0.683 in the testing set. The XGBoost model was interpreted globally and locally using the SHAP method. Na+, Hb, CRP, and globulin levels were independently associated with IVIG-resistance. Our findings demonstrate that ML models can reliably predict IVIG-resistance. Moreover, use of the SHAP method to interpret the established XGBoost model's findings would provide evidence of IVIG-resistance and guide the individualized treatment of Kawasaki disease.

Interpretable machine learning for dermatological disease detection: Bridging the gap between accuracy and explainability

Research on disease detection by leveraging machine learning techniques has been under significant focus. The use of machine learning techniques is important to detect critical diseases promptly and provide the appropriate treatment. Disease detection is a vital and sensitive task and while machine learning models may provide a robust solution, they can come across as complex and unintuitive. Therefore, it is important to gauge a better understanding of the predictions and trust the results. This paper takes up the crucial task of skin disease detection and introduces a hybrid machine learning model combining SVM and XGBoost for the detection task. The proposed model outperformed the existing machine learning models — Support Vector Machine (SVM), decision tree, and XGBoost with an accuracy of 99.26%. The increased accuracy is essential for detecting skin disease due to the similarity in the symptoms which make it challenging to differentiate between the different conditions. In order to foster trust and gain insights into the results we turn to the promising field of Explainable Artificial Intelligence (XAI). We explore two such frameworks for local as well as global explanations for these machine learning models namely, SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME).

Inverse Design of Low-Resistivity Ternary Gold Alloys via Interpretable Machine Learning and Proactive Search Progress.

Ternary gold alloys (TGAs) are highly regarded for their excellent electrical properties. Electrical resistivity is a crucial indicator for evaluating the electrical performance of TGAs. To explore new promising TGAs with lower resistivity, we developed a reverse design approach integrating machine learning techniques and proactive searching progress (PSP) method. Compared with other models, the support vector regression (SVR) was determined to be the most optimal model for resistivity prediction. The training and test sets yielded R2 values of 0.73 and 0.77, respectively. The model interpretation indicated that lower electrical resistivity was associated with the following conditions: a van der Waals Radius (Vrt) of 0, a Vr (another van der Waals Radius) of less than 217, and a mass attenuation coefficient of MoKα (Macm) greater than 77.5 cm2g-1. Applying the PSP method, we successfully identified eight candidates whose resistivity was lower than that of the sample with the lowest resistivity in the dataset by more than 53-60%, e.g., Au1.000Cu4.406Pt1.833 and Au1.000Pt2.232In1.502. Finally, the candidates were validated to possess low resistivity through the pattern recognition method.

An Interpretable Light Attention–Convolution–Gate Recurrent Unit Architecture for the Highly Accurate Modeling of Actual Chemical Dynamic Processes

To equip data-driven dynamic chemical process models with strong interpretability, we develop a light attention–convolution–gate recurrent unit (LACG) architecture with three sub-modules—a basic module, a brand-new light attention module, and a residue module—that are specially designed to learn the general dynamic behavior, transient disturbances, and other input factors of chemical processes, respectively. Combined with a hyperparameter optimization framework, Optuna, the effectiveness of the proposed LACG is tested by distributed control system data-driven modeling experiments on the discharge flowrate of an actual deethanization process. The LACG model provides significant advantages in prediction accuracy and model generalization compared with other models, including the feedforward neural network, convolution neural network, long short-term memory (LSTM), and attention-LSTM. Moreover, compared with the simulation results of a deethanization model built using Aspen Plus Dynamics V12.1, the LACG parameters are demonstrated to be interpretable, and more details on the variable interactions can be observed from the model parameters in comparison with the traditional interpretable model attention-LSTM. This contribution enriches interpretable machine learning knowledge and provides a reliable method with high accuracy for actual chemical process modeling, paving a route to intelligent manufacturing.