Machine learning models for materials
|Machine learning models for materials
|Paul Erhart <firstname.lastname@example.org>
|Chalmers tekniska högskola
|2024-01-29 – 2024-08-01
The Computational Materials Group (https://materialsmodeling.org) studies materials for energy conversion and storage using various modeling techniques, with the goal of improving existing and finding new materials. Part of our work involves the construction and sampling of machine learning (ML) models, including the development of deep neural network models for improving optical sensors as well as the development of so-called neuroevolution potential (NEP) models for atomic scale simulations. Here, we apply for GPU resources for these two applications as described below.
## Hydrogen sensing
The ability to rapidly and reliably detect hydrogen leaks is fundamental for H2 technologies, which are an important part of the sustainable energy economy. Here, Pd-nanoalloy based sensors provide a very promising platform that is, however, still limited in terms of, e.g., response times, tolerance to poisoning agents as well as ageing. In this context we have recently developed a transformer-based deep neural network model (LEMAS) that yields a critical speed-up especially in the technologically important region of low partial pressures (https://arxiv.org/abs/2312.15372). We now want to continue this development in order to not only accelerate H2 sensing but dramatically improve tolerance to poisoning agents and extend their functionality to other sensing modalities (supported by a Vinnova grant). To this end, we build models based on Tensorflow and PyTorch.
## NEP models
The NEP framework, which we co-develop, allows one to build ML models for atomic scale systems that are at the cutting edge of such models in terms of the balance between accuracy, speed, and data efficiency. These models can be used to represent both scalar and tensorial properties, enabling simulations of, e.g., atomic dynamics, chemical reactions with optical or electronic coupling as well as infrared and Raman spectra (https://arxiv.org/abs/2312.05233). Here, we apply for GPU resources for the construction and to a lesser extent inference of these models. During the next year we plan to construct models for, e.g., halide perovskites, metallic alloys, metal carbides, and organic materials (supported by grants from VR, SSF, and KAW). To this end, we will use the GPUMD package (https://gpumd.org), which provides a highly optimized CUDA implementation of the NEP framework.