Source-linked AI summary
Generative adversarial networks (GAN) based efficient sampling of chemical space for inverse design of inorganic materials
Yabo Dan, Yong Zhao, Xiang Li, Shaobo Li, Ming Hu, Jianjun Hu
TL;DR
Efficiently searching the immense inorganic materials design space is difficult because exhaustive simulations and experiments cannot cover it. The paper introduces MatGAN, a GAN trained on known materials to generate hypothetical compounds, with up to 84.5% of ICSD-trained outputs satisfying balanced electronegativity despite no explicit chemical-rule enforcement.
Problem
The immense inorganic chemical space is difficult to search efficiently because brute-force simulations and first-principles methods are too expensive for large-scale screening.
Method
MatGAN represents inorganic materials as sparse 8×85 matrices and uses a deep GAN to learn implicit composition rules from materials databases.
Results
84.5% of GAN-generated samples satisfy balanced electronegativity when trained on charge-neutral and electronegativity-balanced ICSD materials, without explicitly enforcing these rules.
Takeaways & Limitations
The generative models can expand inorganic materials databases with hypothetical candidates for more efficient high-throughput computational screening than exhaustive enumeration.
Takeaways & Limitations
The current representation considers only integer element ratios and therefore does not cover common doped materials with fractional compositions.
Abstract
from arXiv · showhide
A major challenge in materials design is how to efficiently search the vast chemical design space to find the materials with desired properties. One effective strategy is to develop sampling algorithms that can exploit both explicit chemical knowledge and implicit composition rules embodied in the large materials database. Here, we propose a generative machine learning model (MatGAN) based on a generative adversarial network (GAN) for efficient generation of new hypothetical inorganic materials. Trained with materials from the ICSD database, our GAN model can generate hypothetical materials not existing in the training dataset, reaching a novelty of 92.53% when generating 2 million samples. The percentage of chemically valid (charge neutral and electronegativity balanced) samples out of all generated ones reaches 84.5% by our GAN when trained with materials from ICSD even though no such chemical rules are explicitly enforced in our GAN model, indicating its capability to learn implicit chemical composition rules. Our algorithm could be used to speed up inverse design or computational screening of inorganic materials.
1. Introduction
Materials discovery requires efficient methods for searching a vast chemical space because brute-force simulations and first-principles calculations are too expensive at large scale. Generative models offer a way to learn implicit composition rules from existing materials and generate new candidates.
- Large-scale inorganic materials screening is difficult because the chemical design space is vast and computationally expensive to search exhaustively.Brute-force molecular simulations and first-principles methods are described as too expensive for large-scale screening.
- Generative models have been applied to inverse design by learning composition rules from large datasets and generating valid, novel hypothetical materials.Examples include autoencoders, recurrent neural networks, and GANs applied primarily to organic materials.
- MatGAN is proposed as a GAN for efficient sampling of inorganic materials design space using implicit chemical composition rules learned from known materials.The model is trained on inorganic materials databases and does not explicitly specify the chemical rules.
- 84.5% of samples generated by GANs trained on charge-neutral and electronegativity-balanced ICSD materials reproduce those chemical constraints.This result is reported despite the rules not being explicitly enforced during GAN training.
2. Results
MatGAN represents inorganic compositions as sparse matrices and uses adversarial neural networks to generate hypothetical materials. Across validity, stability, uniqueness, novelty, and property-targeted generation tests, performance reflects learned composition patterns but remains limited for underrepresented material categories.
- MatGAN representation: Each material is encoded as an 8×85 sparse 0/1 matrix, with columns representing elements and one-hot atom counts.The representation covers 85 elements, with each element usually appearing fewer than eight times in a compound formula.
- GAN generation model: MatGAN uses a generator and discriminator trained adversarially to produce samples resembling real materials.The generator uses one fully connected and seven deconvolution layers, while the discriminator uses seven convolution and one fully connected layer.
- Chemical-space coverage: 1.69 million charge-neutral and electronegativity-balanced samples remained after filtering 2 million GAN-ICSD outputs.The generated samples occupied a broader region than the ICSD training samples in t-SNE space, potentially expanding the database’s design-space coverage.
- Validity: 92.1% of ICSD-filter generated samples were charge-neutral and 84.5% were electronegativity-balanced without explicitly enforcing those rules.Across binary, ternary, and quaternary enumeration, 62.24% of GAN samples satisfied both rules versus 0.78% for exhaustive enumeration.
- Uniqueness and novelty: At 340,000 generated samples, uniqueness was 68.09% for GAN-OQMD, 85.90% for GAN-MP, and 73.06% for GAN-ICSD.The uniqueness curves declined as sample count increased, with GAN-MP maintaining the highest curve because of training-set composition-distribution differences.
- Targeted generation: GAN-Bandgap generated hypothetical high-bandgap materials with a bandgap distribution similar to its training set.The model was trained on positive-bandgap materials and evaluated using a GBDT predictor with Magpie features.
- Limitations: Nearly 95% of non-decodable OQMD materials were not generated after 2 million GAN-OQMD samples, while 60.26% of decodable samples were rediscovered.Decodable and non-decodable materials also showed distinct element distributions, suggesting that special composition rules or additional model capacity may be needed.
3. Discussion
The inorganic materials design space is too large for exhaustive computation, while existing databases cover only a small fraction of it. The proposed GAN-based approach expands these databases for more efficient screening, but its current representation excludes fractional compositions and crystal-structure information.
- More than 10^12 four-component combinations make the inorganic materials design space impractical for high-throughput experiments or first-principles computations.
- GAN-generated expansions of ICSD, Materials Project, and OQMD could support higher-efficiency computational screening than exhaustive evaluation of billions of candidates.
- Explicit charge-neutrality and electronegativity-balance filters remain too loose for efficient sampling, especially as the number of elements increases.
- The current representation models only integer elemental ratios, excluding common doped materials with fractional compositions.
- The current GAN models do not provide crystal structures such as lattice constants, space groups, or atomic positions.
4. Methods
The methods use filtered inorganic-material datasets and train GANs under dataset-specific settings, with a separate model for band-gap prediction. The primary OQMD training set contains 291,884 compounds after stability and composition-based filtering.
- 4.1 Datasets: 291,884 compounds comprise the final OQMD_L dataset after selecting lowest formation energies, removing single-element compounds, and excluding formation-energy outliers.
- 4.1 Datasets: 63,922 MP compounds and 28,137 ICSD compounds remain after removing single-atom compounds, compounds exceeding eight atoms of any element, and Kr- or He-containing compounds.
- 4.2 GAN Neural networks training: Wasserstein GANs are trained for 1000 epochs with Adam, using learning rates of 0.001 for generators and 0.01 for discriminators.
- 4. Methods: A GBDT band-gap model is trained on 30,186 MP materials with positive band gaps using Magpie features and specified tree and sampling hyperparameters.
- 4. Methods: A supplementary video documenting the work was published on October 22, 2019.