M2Hub - M2 Hub - Unlocking the Potential of Machine Learning for Materials Discovery

Aug 2, 2026

Upload by Patrick

M2Hub is a toolkit designed to accelerate machine learning for materials discovery by providing an integrated workflow rather than isolated datasets. It addresses the gap in modeling materials structures with a platform that delivers access to tasks, datasets, methods, evaluations, and benchmark results. The first release targets virtual screening, inverse design, and molecular simulation, covering 9 datasets for 6 materials types, 56 tasks, and 8 property categories. Benchmarks include realistic out-of-distribution splits and generative-focused resources, with publicly released code.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages18

Reading Time45 min

AccessFree to download

Abstract

Introduction

Background: machine learning in chemical sciences

Need for a unified materials discovery platform

M2Hub overview and benchmark design

Key tasks: virtual screening, inverse design, simulation

Datasets and material properties

Realistic data splits and evaluation

Generative design and benchmark methods

arXiv :2307 .05378v1 [ cond-mat .mtrl-sci ] 14 Jun 2023 M2 Hub: Unlocking the Potential of Machine Learning for Materials Discovery Yuanqi Du1,* Yingheng Wang1,* Yining Huang2 Jianan Canal Li3 Yanqiao Zhu4 Tian Xie5 Chenru Duan6 John M. Gregoire7 Carla P. Gomes1 1 Cornell 2 Northwestern 3 UCB 4 UCLA 5 MSR AI4Science 6 Microsoft Quantum 7 Caltech * Equal Contribution Abstract We introduce M2Hub, a toolkit for advancing machine learning in materials discovery. Machine learning has achieved remarkable progress in modeling molecular structures, especially biomolecules for drug discovery. However, the development of machine learning approaches for modeling materials structures lag behind, which is partly due to the lack of an integrated platform that enables access to diverse tasks for materials discovery. To bridge this gap, M2Hub will enable easy access to materials discovery tasks, datasets, machine learning methods, evaluations, and benchmark results that cover the entire workﬂow. Speciﬁcally, the ﬁrst release of M2Hub focuses on three key stages in materials discovery: virtual screening, inverse design, and molecular simulation, including 9 datasets that covers 6 types of materials with 56 tasks across 8 types of material properties. We further provide 2 synthetic datasets for the purpose of generative tasks on materials. In addition to random data splits, we also provide 3 additional data partitions to reﬂect the real-world materials discovery scenarios. State-of-the-art machine learning methods (including those are suitable for materials structures but never compared in the literature) are benchmarked on representative tasks. Our codes and library are publicly available at [https://github.com/yuanqidu/M2Hub](https://github.com/yuanqidu/M2Hub). 1 Introduction With the methodological advancements in machine learning, an increasing number of machine learning models have been developed and applied to solve scientiﬁc problems, from simulating molecular systems with millions of particles to predicting accurate protein structures Zhang et al.[2018], Jumper et al. [2021] . The primary focus of machine learning in the chemical sciences has remained in the domain of molecular structures,(bio)molecules including small molecules, proteins, RNAs, etc. Atz et al. [2021], Rives et al. [2021], Townshend et al. [2021] . However, materials constitute a large portion of the chemical space which have been signiﬁcantly less studied, especially in the machine learning community. Among scientiﬁc problems, materials discovery plays a vital role in driving innovations and progress across various ﬁelds spanning energy, electronics, healthcare, and sustainability Sanchez-Lengeling and Aspuru-Guzik [2018], Gomes et al. [2021] . However, the traditional trial-and-error approach to materials discovery is expensive and time-consuming. Over decades, classical machine learning methods have already been widely applied in assisting materials discovery, Schmidt et al. [2019] yet the impact of machine learning for solid state materials lags behind its efﬁcacy in other areas of chemical science. Witnessing the success of machine learning in solving grand challenges in science Wang et al. [2018], Jumper et al. [2021], one of the key ingredients is the infrastructure that supports the machine learning community to build the machine learning workﬂow: data preparation/processing, model development, performance evaluation, and model improvement based on the evaluation feedback. While effort Preprint. Under review. Figure 1: M2Hub: Materials discovery meets Machine learning. A-F on the left ﬁgure demonstrates machine learning approaches used in each stage of the materials discovery pipeline on the right ﬁgure (dashed lines denote currently unavailable experiment-related tasks) . has been made to make materials datasets available to the machine learning community Blaisziket al. [2019], Dunn et al. [2020], Clement et al. [2020], Qayyum et a

FAQs

What problem does M2Hub aim to solve in machine learning for materials discovery?

M2Hub targets the lack of an integrated platform for materials discovery, which has slowed development of ML methods for modeling materials structures. It centralizes access to tasks, datasets, methods, evaluations, and benchmarks across the full workflow.

What stages does the first release of M2Hub focus on?

It focuses on three key stages: virtual screening, inverse design, and molecular simulation. These are supported through ML formulations for representations, force fields, and generative materials design.

What does the M2Hub benchmark include in terms of datasets and tasks?

The benchmark is built on curated datasets totaling 9 datasets across 6 types of materials, with 56 tasks spanning 8 material property types. It also adds realistic out-of-distribution data partitions to reflect real-world discovery scenarios.

Aug 2, 2026

Upload by Patrick