Research

My research sits at the intersection of AI for science, interpretable machine learning, and high-performance scientific computing. I build models, algorithms, and software that accelerate response to emerging threats, recover transparent structure from complex data, and power simulation-driven digital twins for healthcare and engineering.

AI for Science

I work in interdisciplinary teams to apply AI to accelerate scientific discovery and engineering design.

Therapeutic Design and Biodefense Preparedness

Why it matters: national readiness for emerging biothreats requires a capabilities-based framework — the infrastructure, models, and pipelines that enable a rapid response to a pathogen we have not yet seen.

My work contributes ML capabilities to that framework: accelerating drug development, predicting and pre-empting viral escape, and shrinking the design–test–iterate loop for therapeutics. Antibody engineering is one concrete instance; the underlying methods (sequence modeling, alignment, combinatorial optimization, library design under constraints) transfer to broader therapeutic and mission-critical problems. Some related work includes:

ProtLib-Designer Pipeline Overview

Schematic overview of the ProtLib‑Designer pipeline for antibody library design. (From the ProtLib‑Designer paper)

Agentic AI for Scientific Discovery

Why it matters: end-to-end LLM agents are unreliable in regulated, mission-critical settings. I build hybrid pipelines where the LLM is scoped to where it is reliable and a deterministic component (solver, fixed evaluation protocol, calibrated threshold) carries the load — producing systems that are auditable, overridable by a domain expert, and deployable where the cost of a wrong answer is high.

Three recent (2026) systems share this design: an LLM that emits structured-JSON priors to an ILP solver for library design under constraints (CIBB ‘26); a Claude-Code-driven autoresearch loop that discovered an interpretable filter rule and improved Leave-One-System-Out ROC-AUC from 0.64 to 0.81 with no LLM at deployment (bioRxiv ‘26); and a prompt-design benchmark with structured confidence signals (LLNL ‘26). The patterns — schema-enforced JSON, deterministic backends, audit trails — transfer beyond protein design to any setting where LLM outputs must be checkable.

CIBB '26: LLM-Guided ILP for Antibody Library Design
bioRxiv '26: Autoresearch Discovery of Interpretable Filter Rules
LLNL '26: Benchmarking LLM Prompting with AF3 Confidence Signals

Autoresearch loop for rule discovery

Autoresearch loop for rule discovery. Candidate-level structural and sequence scores define the available signals; each iteration proposes a rule variant, evaluates it under the fixed LOSO protocol, logs the result, and uses the running-best rule to guide the next proposal.

Cardiac AI and Non-Invasive Imaging

Why it matters: enable non-invasive electroanatomical reconstruction from standard ECGs for better patient stratification.

I develop end-to-end ML pipelines that turn standard clinical data into actionable insights. For example, using only the standard 12-lead surface electrocardiogram, I trained deep neural networks on a dataset of 16,000+ simulated hearts to reconstruct high-resolution electroanatomical maps—potentially enabling non-invasive patient stratification previously only possible via invasive procedures. Some related work includes:

Cardiac ML

Machine learning–driven electroanatomical mapping from 12-lead ECGs. Trained on 16,000+ simulated hearts to reconstruct high-resolution cardiac structure non-invasively. (IEEE 2023 study)

Interpretability in AI

I build transparent AI models that deliver clear, auditable decisions—crucial for healthcare, finance, and regulatory environments.

Symbolic Regression and Scientific Discovery

Why it matters: recover compact, human-readable equations that support scientific discovery and extrapolate beyond training data.

I work with physicists, engineers, and biologists to extract simple, interpretable models from complex data, enabling scientific discovery and deeper understanding of underlying phenomena. Some of my work includes:

Symbolic Regression

Symbolic regression using Deep Symbolic Optimization (DSO) accurately rediscovers the harmonic series identity $H_n = \sum_{k=1}^n \frac{1}{k}$ and Euler’s constant $\gamma$, outperforming neural networks in extrapolation accuracy. (From NeurIPS ’22 work on interpretable AI)

Interpretable Control and Decision-Making

Why it matters: make high-performing control policies auditable in safety-critical settings.

I build interpretable models that govern decision-making in complex systems, ensuring transparency and accountability. In particular, I develop decision-tree policies for reinforcement learning tasks that rival neural nets in performance while being fully transparent. Some of my work includes:

Decision-Tree Controller

Symbolic equations and decision-tree policies governing diverse reinforcement learning environments—offering transparent, high-performing alternatives to neural policies. (From AAAI ’25, ICML ’21, and GECCO ’22 work on interpretable RL)

Scientific Computing

Why it matters: build high-fidelity simulation tools and digital twins for personalized medicine and engineering.

I conduct first-principles simulations of complex multiphysics systems, using continuum mechanics, numerical methods, and high-performance computing. I focus on optimizing both accuracy and throughput on modern HPC platforms. Some of my work includes:

Fluid-Structure Interaction

Simulation of a fluid–structure interaction (FSI) benchmark. The multiphysics model accurately reproduces experimental system dynamics using high-performance finite element solvers. (From NMBI ’18 and CMAME ’16 work on continuum mechanics and FSI)

By integrating these solvers into ML workflows, I aim to power digital twins for personalized medicine, predictive maintenance in robotics, and beyond—delivering end-to-end solutions that marry physical fidelity with data-driven speed.

Digital Twin Heart

Digital twin of the human heart: high-fidelity simulation used to predict disease onset and guide personalized treatment strategies. (Part of ongoing work integrating physics-based solvers with ML pipelines)