Ph.D. Candidate, Mechanical Engineering · Massachusetts Institute of Technology
I research at the intersection of physical AI, thermal fluids modeling, and inverse optics sensing. By combining data-driven methods with high-performance computing, I have been able to construct simulation and inference systems that recover what cannot be measured directly and run them at the scale real problems demand.
Former Google Research simulation researcher · ESSCI Research Fellow (2026–2027) · GE Vernova Fellow (2025–2026)
Three threads run through my research — physical AI, thermal-fluid modeling, and inverse optics sensing — held together by high-performance computing. Most projects sit in more than one.
A differentiable-rendering framework for gradient-based 3-D reconstruction of temperature and species fields from hyperspectral images, built on GPU-accelerated non-gray radiative transfer in CUDA, JAX, and Julia, with diffusion and attention priors for sparse-view reconstruction.
PDE-constrained sensitivity methods that infer material properties and chemical kinetics from thermal imaging, thermocouple, and spectrometer measurements of propagating reaction fronts. Differentiable solvers with adjoint sensitivities make the estimation scalable, and Bayesian inference over reaction-diffusion models keeps results physically consistent and uncertainty-aware.
A global–local attention framework that fuses a learned global latent state with local sensor observations to reconstruct and forecast physical fields — including components no sensor observes — with quantified uncertainty with generative flow matching, trained on reacting-flow LES generated in OpenFOAM with detailed chemistry.
Reinforcement learning combined with greedy active learning to decide what measurement to take next, generating custom wavelength-modulation spectroscopy waveforms at roughly 1 ms latency and accelerating information gain 20×–100× over state-of-the-art methods.
Led development of a CUDA 3-D forward ray-tracing solver with a line-by-line spectral database, coupled to OpenFOAM. BVH-based parallelization gave a >400× speedup and scaled to 1,500 GPUs on a top-3 supercomputer; later work with Kokkos and MPI addressed the communication bottlenecks.
A scalable multi-TPU Lagrangian particle-tracking solver for differentiable wildfire CFD in Swirl-LM, tested to 250 TPU cores. Optimizing SPMD execution meant reading TensorFlow and XLA HLO IR directly — which also surfaced a compiler bug I reported.
Methods and software architecture are inseparable in my work: a model has to be both scientifically useful and practical at accelerator scale.
Differentiable PDE, radiative-transfer, spectroscopy, and rendering models with adjoint sensitivities for gradient-based inference, instrument calibration, and system identification.
Attention and diffusion models, generative priors, and reinforcement learning for reconstruction, forecasting, and experiment design under sparse observation.
Scientific software from custom GPU kernels to multi-node communication design, distributed training, and compiler-level performance investigation.
High-fidelity modeling of fluid flow, combustion, radiative transfer, heat transfer, and multiphase transport, plus the reduced-order models that make them tractable.
A focused selection spanning differentiable inference, optical diagnostics and tomography, scientific machine learning, and accelerated thermal-fluid modeling.
Tricard, N. & Deng, S. High Performance Software Foundation Convention, Kokkos user group meeting, Chicago, IL, March 19. Watch the talk ↗
Tricard, N., Shanbhogue, S., Cherry, M., Chen, Z., Guerra-Garcia, C., Ghoniem, A., & Deng, S. Proceedings of the Combustion Institute.
Tricard, N., Chen, Z., & Deng, S. arXiv:2603.27832. arXiv ↗
Wang, L., Chen, J., Tricard, N., Chen, Z., & Deng, S. arXiv:2603.26023. arXiv ↗
Tricard, N. & Bojko, B. Journal of Propulsion and Power 42(2), 318–332. DOI ↗
Chen, Z., Tricard, N., & Deng, S. Proceedings of the Combustion Institute 41, 105811. DOI ↗
Wang, L., Tricard, N., Chen, Z., & Deng, S. Nanoscale 17, 11812–11863. DOI ↗
Tricard, N., Fraga, G., & Zhao, X. Proceedings of the Combustion Institute 40(1), 105293. DOI ↗
Research roles spanning mathematical formulation, scientific software, performance engineering, and experimental validation.
Differentiable simulation, PDE-constrained inverse modeling, and uncertainty-aware learned models for reacting, radiative, and thermal-fluid systems, validated against optical and thermal diagnostics.
Built distributed multi-TPU Lagrangian particle tracking for differentiable wildfire CFD within Swirl-LM.
Contributed boundary conditions and reduced-order models to a CUDA-accelerated discontinuous-Galerkin Navier–Stokes solver for compressible reacting flow.
Led development of multi-GPU Monte Carlo radiation solvers with line-by-line spectral databases, coupled with CFD for fire, gas-turbine, and high-enthalpy propulsion applications.
Developed 1-D multiphase detonation models reaching a 10⁶× speedup over direct numerical simulation while preserving the key physical trends.
Ph.D., Mechanical Engineering and Computational Science (expected May 2027).
M.S., Mechanical Engineering (Research Assistant). B.S., Mechanical Engineering, minor in Computer Science (Merit Scholar).
Competitive award for collaborative work in flame tomography.
Merit-based full Ph.D. funding; eight awarded across MIT annually.
Merit-based full-funding offer, MIT.
I am a Ph.D. candidate in Mechanical Engineering and Computational Science at MIT, finishing in 2027. My research sits between three fields that increasingly need each other: physical AI, thermal-fluid modeling, and inverse optics sensing.
In practice that means building the solver and the inference method together — CUDA and JAX radiative transfer, adjoint-based PDE-constrained estimation, attention and diffusion models for sparse reconstruction — and then making the whole thing run at the scale the problem actually demands, whether that is 1,500 GPUs on a leadership-class machine or 250 TPU cores at Google Research. Much of it is anchored by real measurements: hyperspectral images, laser spectra, thermocouples and thermal cameras.
The applications have been combustion, fire, propulsion, and materials processing, but the methods are general: recovering what you cannot measure directly from what you can.