MaterIA

A lightweight molecular transformer that autonomously learns to select physics-informed tools for molecular property prediction.

Read the Paper

Understanding matter
at the molecular level

Predicting the properties of molecules — how they interact, store energy, conduct electricity — is fundamental to designing new drugs, advanced materials, and next-generation batteries.

Traditional quantum chemistry methods like Density Functional Theory provide high accuracy but at enormous computational cost, making large-scale molecular screening impractical.

Machine learning offers an alternative, but most models either discard physical knowledge entirely or hard-code it into rigid architectures, limiting adaptability.

Drug Discovery

Screening millions of candidate molecules to identify promising drug compounds requires fast, accurate property prediction at scale.

Advanced Materials

Designing materials with specific electronic, thermal, or mechanical properties demands understanding molecular behavior from first principles.

Energy Storage

Developing better batteries and catalysts depends on predicting how molecules store and transfer energy at the quantum level.

Learnable Tool Injection

Instead of hard-coding physical knowledge into the architecture, MaterIA offers physics-informed tools to the attention mechanism as optional biases. The model learns — without explicit supervision — which tools are relevant for each property and which to ignore.

This behavior, which we call emergent tool discrimination, arises naturally from training. No reward signals. No architectural constraints. The model discovers the optimal tool configuration on its own.

Step 01

Raw Atomic Input

The model receives only atomic numbers and 3D coordinates. No pre-computed features. No molecular fingerprints. Raw data.

Step 02

Tools Are Offered

Physics-informed tools — encoding spatial, electronic, and angular relationships — are injected as learnable biases into the attention mechanism.

Step 03

The Model Decides

Through training, the model autonomously learns to amplify useful tools and suppress irrelevant ones. Each property gets its own optimal tool configuration.

Published and verifiable

Our work is published on ChemRxiv, the preprint server of the American Chemical Society. The results are fully documented and open for peer review.

ChemRxiv

MaterIA: Learnable Tool Injection for Molecular Transformers

Authors: J.E. Peña & Adolfo Romero-Galarza
Platform: ChemRxiv (American Chemical Society)
License: CC-BY-NC-ND 4.0

Access on ChemRxiv

DOI: 10.26434/chemrxiv.15000519/v1

AiGO México

AiGO México is an artificial intelligence research and development company based in San Andrés Cholula, Puebla, México. We build AI systems for scientific discovery and industrial applications.

MaterIA represents our commitment to advancing molecular intelligence — creating models that don't just process data, but develop an understanding of the physical world through learning.

J.E. Peña

Founder & Director General

Location

San Andrés Cholula, Puebla, México