Aionics Releases Advanced Molecular Property Prediction Models for Next Generation Energy Technologies

Aionics is excited to announce today’s launch of its state-of-the-art Molecular Property Prediction Models on Google Cloud. These models, powered by graph neural networks (GNNs), represent a leap forward in materials research and development, and support generative AI-based molecular design processes for energy technologies.

The models are completely general – able to process any arbitrary molecule, including those currently unknown to us – and offer unprecedented accuracy and speed. The integration with Google Cloud means that scientists, engineers, and researchers can now easily predict six key physico-chemical properties: flash point, melting point, boiling point, dipole moment, highest occupied molecular orbital (HOMO) and lowest unoccupied molecular orbital (LUMO). 

These proprietary models are built by Aionics utilizing methodology developed in Prof. Venkat Viswanathan’s lab at Carnegie Mellon University as part of the ARPA-E DIFFERENTIATE program. The intellectual property from CMU is exclusively licensed to Aionics and represents one of the most advanced molecular property optimization capabilities for electrolyte design. 

The Power of Graph Neural Networks

Unlike traditional feature-based models, Aionics models leverage the advanced capabilities of graph neural networks (GNNs), which learn directly from molecular structures represented as graphs. This approach enables the models to uncover complex patterns within molecular data, resulting in more accurate predictions of properties. The ability to extract insights directly from molecular graphs also enhances prediction reliability, making these models more powerful than conventional approaches. 

The training datasets underlying the Aionics GNN models come from a proprietary combination of atomistic simulations and experimental data. The volume of available data on these physico-chemical properties is increasing rapidly, and GNNs are able to more efficiently learn the underlying structure-property relationships in large datasets than feature-based methods. 

Why Molecular Properties Matter

Molecular properties play a crucial role in the design of materials for a wide range of energy applications. For instance, designing battery electrolytes requires satisfying key molecular characteristics: the temperature range of the liquid phase (i.e. melting and boiling points), oxidative stability (highest occupied molecular orbital) and reductive stability (lowest unoccupied molecular orbital), salt solubility (dipole moment), and flammability (flash point).

Given the importance of these six criteria, Aionics has deployed these models first so researchers can quickly predict these properties and make data-driven decisions with confidence.

Model Usage and Capabilities

The models are able to predict on any SMILES string—a simplified textual representation of a molecule—as input, and outputs a predicted property value and an estimate of the standard deviation, providing users with a confidence measure for the prediction. Generative AI offers the potential to generate novel molecules with a valid SMILES string but unknown properties; users can utilize these models to rapidly and reliably predict the properties for these novel molecules.

Aionics has been a part of the Google for Startups Cloud Program. This has provided us access to developer-friendly technology, resources, enabling us to offer this molecular property prediction service.  We are able to offer reliable, fast, and highly scalable predictions; users can deploy these models to score massive datasets such as the entire ZINC catalog of commercially available molecules. Whether you’re working in academia, industry, or R&D, our models can help you streamline your workflow and accelerate your material discovery and design process.

Get Started Now

Ready to try it out? Submit a request below to get started and explore how Aionics Molecular Property Prediction Models can help drive your research forward.