Energy efficiency and trustworthiness are reshaping how scientists use artificial intelligence to study molecular systems. Pratyush Tiwary, a chemistry professor at the University of Maryland Institute for Health Computing, received a $902,000 grant from the U.S. Department of Energy to develop leaner, physics-informed AI tools that can predict rare chemical events like crystal formation without consuming massive amounts of computing resources.

The three-year award marks the second renewal of a project that began in 2020, bringing total Department of Energy support to $2.3 million over nine years. The 30% budget increase signals growing recognition that the next generation of AI for chemistry must solve a critical problem: today's AI models are brittle and inflexible, working reliably only under the exact conditions they were trained on.

What Makes Current AI Chemistry Models Fall Short?

Most AI systems trained on chemical data struggle when conditions change. If a model learns to predict molecular behavior at room temperature and standard pressure, moving to a different temperature or pressure often requires collecting new experimental data, running expensive simulations, and building an entirely new model. This limitation makes AI impractical for real-world chemistry, where conditions vary constantly.

"Today's AI chemistry models are often brittle, working only for the temperatures, pressures, chemical types and other factors used in the model's training," explained Tiwary. "In chemistry, moving to a new condition can require new data, expensive simulations and a new model."

Pratyush Tiwary, Professor of Chemistry and Biochemistry at University of Maryland

Tiwary's team has already built AI tools that help computer simulations identify which molecular motions matter most, making it possible to study rare events that would normally take far too long to observe with standard simulations. The new funding will push this work further by embedding thermodynamics and statistical physics directly into AI models, allowing them to learn from less data and predict reliably across conditions they've never encountered.

How to Build AI That Learns Like a Physicist

• Physics-Informed Design: Incorporating thermodynamics and statistical physics into AI models allows them to generalize beyond their training data, similar to how a physicist understands why molecules behave certain ways rather than just memorizing patterns.
• Data and Energy Efficiency: By building fundamental physics principles into the model architecture, researchers can achieve accurate predictions using significantly less training data and computational power than conventional deep learning approaches.
• Explainability and Trust: The team focuses not just on what answer the model produces, but how it learned to produce that answer, making the AI's reasoning transparent and trustworthy for scientific applications.

The research team has already published findings on these methods in peer-reviewed journals. In 2026, they described latent thermodynamic flows in the Royal Society of Chemistry's Chemical Science Journal, and in 2025, they published work on crystal nucleation in the Proceedings of the National Academy of Sciences.

Why Does This Matter for Materials and Drug Discovery?

Tiwary's work addresses a fundamental challenge in computational chemistry and materials science. Scientists need to predict how molecules will behave under new conditions to design better drugs, discover novel materials, and optimize chemical processes. Traditional computer simulations can take weeks or months to model rare events like the initial stages of crystal formation. AI can speed this up dramatically, but only if the models are reliable and don't require retraining every time conditions change.

The broader computing landscape is shifting toward hybrid systems that combine AI, quantum computing, and classical supercomputers to tackle complex scientific problems. Tiwary's energy-efficient AI fits into this emerging ecosystem as a tool that makes scientific discovery faster and more sustainable.

"Building thermodynamics and statistical physics into AI taps into these benefits," Tiwary stated. "And because such gains must be trusted, our process also asks how the model learned, not just what answer it produced. This is part of our group's long-term effort to build robust, trustworthy and energy-efficient AI-powered science."

Pratyush Tiwary, Professor of Chemistry and Biochemistry at University of Maryland

The Department of Energy's Basic Energy Sciences program, which funded this work, supports fundamental research behind energy technologies spanning generation, conversion, transmission, storage, and use. By making AI more efficient and generalizable, Tiwary's team contributes directly to the agency's mission of deploying AI models that enable the future of energy sciences.

As AI becomes more central to scientific discovery, the focus is shifting from simply building larger, more powerful models to building smarter, more efficient ones that can be trusted and understood. Tiwary's work exemplifies this shift, showing that the next frontier in AI-powered chemistry isn't about processing more data, but about embedding scientific wisdom into the models themselves.