AI for materials design: Generative AI with multi-fidelity strategies

The design of new molecules and materials is often hindered by the vastness of chemical and microstructural design spaces and the high cost of obtaining high-fidelity property labels through quantum or physics-based simulations. This dissertation introduces a unified framework that combines generative artificial intelligence (AI), hierarchical transfer learning, multi-fidelity modeling, and graph-driven voxel-based analysis to address four critical challenges in materials discovery: the need for syntactically valid and interpretable generative models, the data inefficiency of high-fidelity property prediction, the unreliability of models under distribution shift, and the lack of scalable tools for characterizing dynamic structural domains. First, we present MolGen-Transformer, a transformer-based molecular language model trained on a dataset of 198 million molecules using the SELFIES representation. It achieves perfect reconstruction accuracy, generates chemically diverse and valid molecules, and offers a compact and interpretable latent space suitable for downstream design tasks. Second, we build on this well-trained model by developing a hierarchical property prediction framework that fuses MolGen-Transformer embeddings with both low- and high-fidelity labels. This multi-fidelity approach reduces the dependence on expensive density functional theory (DFT) data by up to fourfold. It also integrates uncertainty quantification via ensemble modeling to support reliable, property-driven molecular design through latent space path search. Third, we generalize the multi-fidelity paradigm beyond molecules by applying it to microstructure–property prediction for organic photovoltaics. We demonstrate that a combination of learned microstructure embeddings and limited high-fidelity simulations enables accurate prediction of device-level performance with data efficiency. Fourth, we introduce MDVoxelizer, a modular framework that integrates graph-based structural filtering with voxel-based spatial mapping to quantify local crystallinity in molecular dynamics simulations. This enables interpretable, time-resolved, and machine learning–compatible representations of complex, evolving molecular systems. Collectively, these contributions establish a robust, generalizable approach to AI-guided design. By tightly integrating generative modeling, fidelity-aware learning, structural characterization, and uncertainty estimation, this work enables scalable exploration and optimization in chemical and materials spaces while mitigating computational cost and predictive risk.

Name thesis:degree_name

Doctor of Philosophy

Level thesis:degree_level

dissertation

Discipline thesis:degree_discipline

Artificial intelligence

Department dc:contributor.department

Department of Mechanical Engineering

Grantor

Iowa State University

Year dc:date.issued

2025

Author dc:creator

• Yang, Chih Hsuan

Advisors dc:contributor.advisor

• Ganapathysubramanian, Baskar
• Sarkar, Soumik
• Krishnamurthy, Adarsh
• Liu, Hailiang
• Li, Qi

Language dc:language.iso

en

DOI

https://doi.org/10.31274/td-20260223-95

OAI identifier oai:identifier

oai:dr.lib.iastate.edu:20.500.12876/KrZJOYXr