Source-linked AI summary
A General-Purpose Machine Learning Framework for Predicting Properties of Inorganic Materials
Logan Ward, Ankit Agrawal, Alok Choudhary, Christopher Wolverton
TL;DR
Locating novel metallic glass alloys is slow and resource-intensive, motivating predictive models from experimental data. This framework uses chemically diverse attributes, partitions similar materials, and finds decision trees effective for accurate materials-property prediction.
Problem
Conventional methods for locating novel metallic glass alloys are slow and resource-intensive.
Method
The framework uses elemental property statistics and trains separate models on material subsets to improve prediction accuracy.
Results
Decision trees perform best with the attribute set, while dataset partitioning can significantly increase prediction accuracy.
Takeaways & Limitations
High-predictive-accuracy decision trees are excellent choices for materials design.
Takeaways & Limitations
The framework’s decision-tree models are less suited to understanding the underlying mechanisms behind material properties.
Abstract
from arXiv · showhide
A very active area of materials research is to devise methods that use machine learning to automatically extract predictive models from existing materials data. While prior examples have demonstrated successful models for some applications, many more applications exist where machine learning can make a strong impact. To enable faster development of machine-learning-based models for such applications, we have created a framework capable of being applied to a broad range of materials data. Our method works by using a chemically diverse list of attributes, which we demonstrate are suitable for describing a wide variety of properties, and a novel method for partitioning the data set into groups of similar materials in order to boost the predictive accuracy. In this manuscript, we demonstrate how this new method can be used to predict diverse properties of crystalline and amorphous materials, such as band gap energy and glass-forming ability.
2 Department of Electrical Engineering and Computer Science, Northwestern University, · Introduction
The paper introduces a general-purpose machine-learning framework for predicting materials properties from composition using reusable attributes and a novel similarity-based data-partitioning scheme. It demonstrates broad applicability across crystalline and amorphous materials, computational and experimental data, and continuous and discrete properties.
- Introduction: Materials databases have enabled predictive models, but their growing size makes discovering design rules through human intuition impractical.Machine learning offers a way to create accurate models quickly and automatically.
- Introduction: Existing materials machine-learning models cover only a small fraction of relevant properties and datasets.Broadly applicable models were still absent for band gap energy and glass-forming ability despite large databases.
- Introduction: A systematic workflow for transforming raw materials data into quantitative model inputs had not yet emerged.The authors contrast this gap with chemoinformatics, which has flexible molecular-structure descriptors and standard input-selection approaches.
- Introduction: The authors present a general-purpose framework that predicts materials properties from composition using reusable attributes across diverse materials problems.The attributes serve as machine-learning inputs and are intended to support broad reuse.
- Introduction: The framework reduces development of a new property model to selecting a well-performing machine-learning algorithm once flexible inputs are available.Algorithm selection is described as a well-studied data-science problem.
- Introduction: A novel partitioning scheme groups similar materials and trains separate models for each group to enhance prediction accuracy.The method first partitions the data into similar material groups.
- Introduction: The method applies to amorphous or crystalline materials, computational or experimental data, and properties with continuous or discrete values.Its versatility is demonstrated through solar-cell-material prediction from DFT data and metallic-glass-alloy suggestion from experimental glass-forming measurements.
- Introduction: The framework is intended to enable rapid model creation from materials databases and advance rational materials design.The authors frame this as the basis for quickly creating models from available data.
Results and Discussion
The framework uses broad composition-based attributes and material-group partitioning to build accurate models across diverse inorganic-material properties. Results span crystalline-property prediction and metallic-glass discovery, while noting limitations of average-composition descriptions and the need for richer attribute libraries.
- General-purpose attributes: A general-purpose attribute set should capture statistics of many elemental properties because important attributes vary substantially across material properties.Examples include pseudopotential radii, ionization potential, and electron affinity for phase energy differences, versus atomic number, mass, and radii differences for melting temperature.
- Models and partitioning: Decision-tree ensembles performed best with the attribute set, while partitioning materials into similar groups improved prediction accuracy by training separate models on each subset.These methods provide high predictive accuracy but are less interpretable for understanding underlying mechanisms.
- Property-dependent attributes: The best attributes differ by property and compound type, with electronegativity describing crystalline formation energy, p-shell electron fraction and periodic-table row describing band gap, and other statistics distinguishing compound classes.Intermetallic formation energy is associated with melting-temperature variance and d-electron-number differences, whereas nonmetal-containing compounds are associated with mean ionic character.
- Crystalline compounds: Above 0.91 correlation coefficients were obtained between measured and predicted values for all three crystalline properties, with reduced-error pruning decision-tree ensembles achieving the lowest cross-validation mean absolute error.The reported comparison covered 10 machine-learning algorithms.
- Band-gap screening: 67% of predictions had the desired band gap energy after partitioning the dataset into groups of similar compounds, compared with approximately 46% using one model trained on the entire dataset and just over 12% by random selection.The screening list contained 30 selected compounds, and 223 of 4500 predicted ternary compounds were likely to have favorable band gap energies.
Conclusions
The work introduces a general-purpose machine-learning framework for predicting diverse materials properties and demonstrates its applicability to crystalline photovoltaic compounds and metallic glass alloys. The framework uses chemically informed attributes and similarity-based dataset partitioning to create accurate models and support broader, faster deployment in materials design.
- The framework predicts properties across a wide variety of materials, including crystalline electronic properties and metallic-alloy glass formability.
- The demonstrations address discovering potential crystalline photovoltaic compounds and identifying candidate metallic glass alloys.
- Model accuracy is enhanced by chemically informed attributes and partitioning the dataset into groups of similar materials.
- The strategy requires selecting the most accurate machine-learning algorithm and testing partitioning strategies, processes that could eventually be automated.
- The method’s versatility is intended to enable quicker deployment and wider-scale use of machine learning in designing new materials.
Methods
The study used Weka and Magpie to build machine-learning models, compute attributes, validate experiments, and search for new materials.
- Software and data: Weka was used to create all machine-learning models.Weka is available under an open-source license.
- Software and data: Magpie was used to compute attributes, perform validation experiments, and search for new materials.Magpie is available under an open-source license.
- Software and data: The software, training datasets, and input files were provided in the associated Supplementary Information.
Contributions
The project was conceived by CW, who jointly developed the method with LW, AA, and AC. LW implemented the software, performed the calculations, and led manuscript writing, with guidance and contributions from the other authors.
- CW conceived the project and jointly developed the method with LW, AA, and AC.
- LW wrote all software and performed the necessary calculations with help and guidance from AA and AC.
- LW led manuscript writing, with all other authors contributing.
Figures
The figures show that machine-learning strategies improve band-gap target discovery over random selection and predict glass-forming ability across the Al-Ni-Zr ternary. The band-gap models use either a single model or models trained separately on groups of similar materials.
- Band-gap discovery: The single-model band-gap strategy was trained on computed band-gap energies for 22667 ICSD compounds.This strategy used one model rather than separate models for material subsets.
- Band-gap discovery: The second band-gap strategy partitioned the data into groups of similar materials and trained a separate model on each subset.This is the alternative machine-learning strategy compared in Figure 1.
- Band-gap discovery: After 30 guesses, the machine-learning approach found over 5 times more compounds within the desired band-gap range than random selection.Figure 1 compares random selection with two machine-learning strategies for locating compounds with band gaps in a desired range.
- Glass-forming ability: The model predicted relative glass-formation likelihood across the Al-Ni-Zr ternary, with 1 as most likely and 0 as least likely.The predictions were generated without training on measurements from the Al-Ni-Zr ternary or its constituent binaries.
Tables
The tables compare machine-learning algorithms for predicting material properties from OQMD data and present predicted candidates for solar-cell materials and metallic glass alloys. The candidate predictions use models trained on DFT band gaps or experimental glass-forming ability measurements.
- Algorithm comparison: Table 1 compares several machine-learning algorithms using mean absolute error from 10-fold cross-validation on a single model trained with DFT-predicted properties for 228,676 materials.The data are from the OQMD.
- Solar-cell candidates: Table 2 lists ternary compounds predicted as solar-cell candidates, with their nominal compositions and predicted band gap energies.The compounds were predicted using methods developed in Ref. 16, while band gaps were predicted using an OQMD-trained DFT band-gap model.
- Metallic glass candidates: Table 3 lists candidate metallic glass alloys predicted by a model trained on experimental measurements of glass-forming ability.The alloys were selected for their predicted ability to form amorphous ribbons by melt spinning from 24 million candidates.