MIT Postdoc Deng Bowen Joins Fudan to Build AI Scientists for Materials Discovery

Background and Career Trajectory

MIT postdoctoral researcher Deng Bowen (邓博文) will join Fudan University's Artificial Intelligence Innovation and Industry Research Institute (AI³) in early 2027 as an assistant professor and PhD supervisor. He has already begun preparing his own laboratory — the Atomic Intelligence Lab — with a clear mission: build reliable AI scientists to accelerate new materials development, from atomic simulation to experimental validation.

Deng's academic path traces a progression from computational materials science to AI-driven materials research:

Undergraduate at UC Santa Barbara : first-principles calculations on topological Dirac semimetals.

2020: Internship at Microsoft Research Asia introduced him to graph neural networks for materials, prompting a shift toward materials AI.

2021: PhD in Materials Science & Engineering at UC Berkeley ; received the MRS Graduate Student Silver Award .

PhD focus: machine learning simulation of atomic interactions .

Post-PhD: Postdoc at MIT with Professor Rafael Gómez-Bombarelli , extending research from single models to complete scientific workflows.

CHGNet: Efficient Atomic Interaction Modeling

In 2023 , Deng published CHGNet as first author in Nature Machine Intelligence . CHGNet is a machine learning model that learns from large-scale materials computation data to simulate atomic interactions more efficiently, providing a new tool for battery materials and other domains. The paper has garnered over 1,000 citations ; the code is open-source and integrated into community libraries such as MatGL .

Reliability Validation: MatPES and Systematic Bias

Beyond simulation capability, Deng turned to the question of data and model reliability . In the MatPES project, the team filtered ~400,000 structures from ~281 million molecular dynamics snapshots to build a high-quality materials dataset.

He also collaborated on evaluating machine learning predictions of crystal stability , revealing that universal materials models can exhibit systematic biases . Even when predictions appear reasonable, they require rigorous validation before use in scientific research.

Applications to Battery Materials

Deng applied ML models to concrete materials problems, including:

Voltage hysteresis mechanism in disordered rock-salt cathodes.

Lithium-ion transport in halide solid-state electrolytes.

Results published in top journals: JACS , Joule , ACS Energy Letters .

AtomisticSkills: AI Agents for Scientific Workflows

At MIT, Deng participated in AtomisticSkills , which equips AI agents with specialized tools for materials simulation — enabling them to call computational programs, chain research steps, and verify execution results. This shifts AI's role from assisted prediction to active participation in computation and analysis . Ensuring accuracy and reproducibility at each step remains a key challenge.

Atomic Intelligence Lab Vision

The trajectory "simulation capability → reliability validation → agent workflow" directly underpins the Atomic Intelligence Lab's goal: construct reliable AI scientists that can autonomously drive materials discovery from atomic-scale modeling through experimental verification.