MLMD: a programming-free AI platform to predict and design materials

The discovery of novel materials is a cornerstone of technological advancement, impacting fields like aerospace, biomedicine, and energy. However, traditional material research relies heavily on a trial-and-error approach, which is costly and time-consuming. The vast chemical space and intricate relationships between composition, processing, structure, and properties (CPSP) present significant challenges. This research gap underscores the urgent need for efficient material discovery strategies. Artificial intelligence (AI), particularly machine learning (ML), offers a powerful alternative, enabling data-driven approaches to unravel CPSP relationships and predict material properties. Numerous AI toolkits and platforms have been developed for materials science; however, they often suffer from limitations, including a primary focus on property prediction, a lack of user-friendliness for non-programmers, and poor performance with limited data. This paper addresses these shortcomings by introducing MLMD, a novel AI platform designed to facilitate end-to-end materials discovery and design. Unlike existing platforms, MLMD emphasizes a programming-free interface, making it accessible to materials scientists without extensive programming expertise. Furthermore, it integrates multiple AI techniques to handle data scarcity and efficiently explore the vast design space, leading to the discovery of novel materials with superior properties.

Several AI platforms have emerged within the materials and physical computation community, each with its strengths and weaknesses. Materials Cloud focuses on ab initio computations and material similarity, while the Materials Project leverages quantum computations and a large inorganic materials database. AFLW-ML and JARVIS-ML offer crystal property prediction tools based on DFT calculations or ML surrogate models. Matminer and Magpie provide useful ML libraries for materials research. While some general-purpose AI toolkits exist, many require programming skills, posing a barrier for materials scientists without coding experience. Moreover, most platforms concentrate on model construction, neglecting the crucial aspect of inverse materials design – designing materials with specific properties. The current landscape lacks a comprehensive, user-friendly platform that seamlessly integrates various AI techniques for end-to-end materials design.

MLMD is designed as an end-to-end platform for materials design, requiring no programming knowledge. Users upload data in CSV format, specifying features (material components and processes) and target variables (material properties). The core modules of MLMD are: (1) Database: Provides access to material data, including outlier detection functionalities. (2) Data Visualization: Offers initial data overview and distribution analysis. (3) Feature Engineering: Handles missing and duplicate values, assesses feature correlation, and ranks feature importance. (4) Classification: Supports various supervised learning models, hyperparameter tuning, and model ensemble. (5) Surrogate Optimization: Integrates predictive models into optimization algorithms to find optimal material compositions and processes. (6) Active Learning: Addresses data scarcity using Bayesian-based active learning, balancing exploration and exploitation to efficiently discover materials. (7) Interpretability: Provides model explanation using SHAP values. Three primary workflows are supported: model inference, surrogate optimization, and active learning. Model inference uses a trained model to predict properties. Surrogate optimization integrates a predictive model into optimization algorithms to discover materials with targeted properties. Active learning employs a Bayesian sampling strategy to iteratively guide experiments and refine the model. The platform supports various ML algorithms for both classification and regression tasks, enabling users to choose suitable models and automatically optimize hyperparameters. MLMD integrates transfer learning and heuristic algorithms to handle mixed data problems. The platform was tested on diverse datasets including perovskites, steels, and high-entropy alloys, demonstrating its robustness and effectiveness in both model construction and inverse design.

The MLMD platform demonstrated strong performance in both classification and regression tasks. In classification tasks involving polycrystalline ceramics, zinc alloys, and high-entropy alloys, the default MLMD-implemented models achieved over 80% accuracy in 10-fold cross-validation. Hyperparameter tuning further improved accuracy, exceeding the performance of baseline models. In regression tasks predicting the fracture strength of low-alloy steels, Curie temperature of perovskites, and flow stress of superalloys, MLMD-implemented models outperformed baseline models, achieving high R² values (0.9427, 0.8480, and 0.9288 respectively). The surrogate optimization module successfully designed RAFM steels with enhanced strength and ductility. The active learning module efficiently discovered new high-hardness high-entropy alloys, comparable to results achieved in previous studies using iterative loops of AI and knowledge-based methods. The platform's effectiveness was demonstrated across various materials, highlighting its versatility and robustness. The user-friendly, programming-free interface makes it readily accessible to a wider range of materials scientists.

MLMD significantly advances the field of materials informatics by providing a user-friendly, programming-free platform capable of end-to-end materials design. The platform's success in handling both abundant and scarce data, combined with its ability to perform both property prediction and inverse design, addresses critical limitations of existing tools. The successful application of MLMD to various materials systems, including perovskites, steels, and high-entropy alloys, underscores its versatility and potential for broad impact. The integration of active learning effectively tackles the common challenge of limited data in materials science, enabling efficient exploration of the design space and the discovery of novel materials with enhanced properties. The platform's ability to streamline the materials design process has the potential to substantially accelerate the pace of materials innovation, reducing research costs and time.

MLMD presents a significant advancement in AI-driven materials design. Its programming-free interface, coupled with its comprehensive suite of functionalities, makes it a powerful tool for materials scientists of all programming skill levels. The successful application to diverse materials systems and the efficient handling of data scarcity demonstrate its versatility and potential to revolutionize the materials discovery process. Future work will focus on enhancing the platform's capabilities by incorporating more sophisticated ML algorithms and expanding the available functionalities, further solidifying its position as a critical tool for accelerating materials innovation.

While MLMD offers a user-friendly and powerful platform, some limitations exist. The accuracy of the predictions relies heavily on the quality and quantity of the input data. The platform's performance might be affected by the inherent biases or limitations in the underlying datasets. The interpretability of the models, while enhanced by the inclusion of SHAP values, can still be a challenge in understanding the complex relationships between features and properties. Future development should focus on addressing these limitations and enhancing the platform's robustness and explainability.