Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Exploration of Novel Neuromorphic Methodologies for Materials Applications

Derek Gobin, Shay Snyder, Guojing Cong, Shruti R. Kulkarni, Catherine Schuman, Maryam Parsa

TL;DR

This work investigates neuromorphic alternatives to graph neural networks for materials property prediction, focusing on reservoir computing and hyperdimensional computing (GraphHD and SSP-GrapHD) applied to a 54-point subset of the Materials Project for bandgap classification and regression. The study finds that SSP-GrapHD delivers the strongest results, achieving MAE $0.5181$ and accuracy $0.8235$, outperforming the GraphHD and GNN baselines, including ALIGNN. The results demonstrate that hyperdimensional representations can effectively encode molecular graphs with potential advantages for low-power, hardware-friendly implementations. These findings motivate large-scale evaluation on full datasets and exploration of richer atomic features, multi-hop memory, and cross-domain applications in physics and imaging.

Abstract

Many of today's most interesting questions involve understanding and interpreting complex relationships within graph-based structures. For instance, in materials science, predicting material properties often relies on analyzing the intricate network of atomic interactions. Graph neural networks (GNNs) have emerged as a popular approach for these tasks; however, they suffer from limitations such as inefficient hardware utilization and over-smoothing. Recent advancements in neuromorphic computing offer promising solutions to these challenges. In this work, we evaluate two such neuromorphic strategies known as reservoir computing and hyperdimensional computing. We compare the performance of both approaches for bandgap classification and regression using a subset of the Materials Project dataset. Our results indicate recent advances in hyperdimensional computing can be applied effectively to better represent molecular graphs.

Exploration of Novel Neuromorphic Methodologies for Materials Applications

TL;DR

This work investigates neuromorphic alternatives to graph neural networks for materials property prediction, focusing on reservoir computing and hyperdimensional computing (GraphHD and SSP-GrapHD) applied to a 54-point subset of the Materials Project for bandgap classification and regression. The study finds that SSP-GrapHD delivers the strongest results, achieving MAE and accuracy , outperforming the GraphHD and GNN baselines, including ALIGNN. The results demonstrate that hyperdimensional representations can effectively encode molecular graphs with potential advantages for low-power, hardware-friendly implementations. These findings motivate large-scale evaluation on full datasets and exploration of richer atomic features, multi-hop memory, and cross-domain applications in physics and imaging.

Abstract

Many of today's most interesting questions involve understanding and interpreting complex relationships within graph-based structures. For instance, in materials science, predicting material properties often relies on analyzing the intricate network of atomic interactions. Graph neural networks (GNNs) have emerged as a popular approach for these tasks; however, they suffer from limitations such as inefficient hardware utilization and over-smoothing. Recent advancements in neuromorphic computing offer promising solutions to these challenges. In this work, we evaluate two such neuromorphic strategies known as reservoir computing and hyperdimensional computing. We compare the performance of both approaches for bandgap classification and regression using a subset of the Materials Project dataset. Our results indicate recent advances in hyperdimensional computing can be applied effectively to better represent molecular graphs.
Paper Structure (8 sections, 4 equations, 2 figures, 1 table)

This paper contains 8 sections, 4 equations, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Reservoir Computing
  5. Hyperdimensional Computing
  6. Results
  7. Conclusion and Future Work
  8. Acknowledgment

Figures (2)

  • Figure 1: Encoding the structure of (A) PbB2 molecule into (B) three spike vectors.
  • Figure 2: A simplified visualization of encoding (A) parent node A into (B) an object hyperdimensional vector that incorporates neighbor nodes and a spatial hyperdimensional vector that represents position. This process would be repeated for each node in the graph.