Learning surrogate models for computational fluid dynamics
Current Ph.D. research in the SmartDATA Lab with Dr. Joel B. Harley: replacing expensive computational-fluid-dynamics solver runs with fast learned surrogates that are held accountable to the physics they approximate.
Ph.D. research in the SmartDATA Lab: learned surrogates that stand in for expensive CFD solvers, judged on stability, physical consistency, and generalization.
The project asks a simple question with a demanding answer: when can a learned model stand in for a numerical fluid-dynamics solver? A surrogate that is fast but physically wrong is worse than useless in a design loop, so the program treats evaluation — rollout stability, conservation behavior, and out-of-distribution generalization — as a first-class research object rather than an afterthought.
High-fidelity CFD is accurate and expensive; design iteration wants answers in milliseconds. This project studies how much of the solver's job a learned model can take over — and how to tell when it should not be trusted.
Why surrogates: the cost of asking a solver a question
A high-fidelity computational-fluid-dynamics run integrates the governing equations over a fine mesh, and the price is hours of compute per case. That cost is acceptable for a final answer and prohibitive for everything before it: design sweeps, optimization loops, uncertainty quantification, and real-time settings all want thousands of evaluations, not one.
Surrogate modeling replaces most of those evaluations with a learned approximation trained on solver output. The interesting research questions are not whether a network can fit the training snapshots — it can — but whether the surrogate stays stable when rolled forward in time, respects physics it never explicitly saw, and degrades honestly when asked about conditions outside its training distribution.
Four modeling families, one evaluation discipline
The program deliberately spans the main families of learned surrogates rather than committing to one architecture early, because each family encodes a different assumption about which structure in the flow is worth preserving.
Neural operators
FNO- and DeepONet-style models that learn maps between function spaces, so one trained model can answer for new initial and boundary conditions.
Mesh graph networks
Message passing over the simulation mesh itself, keeping unstructured geometry and local interactions explicit.
Physics-informed training
PDE residuals and conservation terms in the loss, so the governing equations constrain the model where data is sparse.
Latent dynamics
Compress the field to a low-dimensional latent state and learn its dynamics — the same state-space discipline as my earlier time-series work.
Figure 1 — Expensive solver runs and physics constraints train a family of surrogates; the fast model is then audited for stability and physical consistency, not just pointwise error.
What counts as success
Pointwise error on held-out snapshots is the weakest claim a surrogate can make. The evaluation questions that matter compound over time and under distribution shift:
Rollout stability — does autoregressive prediction stay bounded and physical over long horizons, or does error accumulate into blow-up?
Physical consistency — are conserved quantities approximately conserved and boundary conditions respected, even though the model is not a solver?
Generalization — what happens on geometries, boundary conditions, and flow regimes the training set never covered?
Honest speed accounting — speedups only mean something when measured at the accuracy actually required, on matched hardware.
The through-line from my earlier work
This direction is a change of domain, not of discipline. The neural-imaging work in CNEL was about extracting trustworthy structure from noisy spatiotemporal data; the experiment-infrastructure work was about making model comparisons fair and repeatable; the state-space projects were about latent dynamics with interpretable failure modes. CFD surrogate modeling stresses all three habits at once: fluid fields are spatiotemporal data with known governing structure, surrogate families must be compared under matched budgets, and a latent-dynamics view is itself one of the candidate architectures.
Where it stands
The work has been under way in the SmartDATA Lab since Fall 2025. Results, code, and write-ups are not yet public; this page will link them as they are released. In the meantime, the CNEL and IEEE SPS project pages are the best public evidence of how I run a research program.
FAQ
What is a CFD surrogate model?
A machine-learning model trained on the outputs of an expensive computational-fluid-dynamics solver so it can approximate new solutions in a fraction of the time. The solver stays the ground truth; the surrogate makes the solver's answers affordable inside design and analysis loops.
Why work across four model families instead of picking one?
Each family bakes in a different assumption: neural operators generalize across conditions, mesh networks respect geometry, physics-informed losses inject the governing equations, and reduced-order models make the latent dynamics explicit. Comparing them under one evaluation discipline is the point of the program.
Is any of this work public yet?
Not yet. The work started in Fall 2025, and results and code will be linked from this page when they are released.
Snapshot
Track
Research · scientific ML · SmartDATA Lab
Status
Active research since Fall 2025 · results not yet public.
