Auto-MatRegressor: liberating machine learning alchemists.

Area of Science:

• Materials Science
• Computational Materials Science
• Data Science

Background:

• Machine learning (ML) is crucial for predicting material properties by identifying patterns and relationships.
• Current ML model development for materials science is hindered by time-consuming and labor-intensive experimental data acquisition.
• High-accuracy ML models require extensive data and expertise in algorithm selection and hyperparameter tuning.

Purpose of the Study:

• To develop an automated machine learning (ML) modeling method for materials property prediction.
• To reduce the computational cost and experimental effort in building accurate ML models for materials science.
• To introduce Auto-MatRegressor, a meta-learning based approach for automating algorithm selection and hyperparameter optimization.

Main Methods:

• Developed Auto-MatRegressor, an automatic modeling framework utilizing meta-learning on historical materials datasets.
• Incorporated 27 meta-features characterizing datasets and the performance of 18 common materials science algorithms.
• Implemented a collaborative meta-learning strategy enhanced with domain knowledge from a materials categories tree for algorithm recommendation.

Main Results:

• Auto-MatRegressor demonstrated efficient algorithm selection and hyperparameter optimization compared to traditional methods.
• The automated approach significantly reduced computational cost while achieving good prediction accuracy across 60 diverse datasets.
• The system showed effectiveness in accelerating the construction of ML models for materials property prediction.

Conclusions:

• Auto-MatRegressor offers a scalable and efficient solution for automated ML model development in materials discovery and design.
• The meta-learning approach effectively leverages past modeling experience to guide new model construction.
• This method supports dynamic expansion, enabling continuous improvement as more data and algorithms become available.