Aditya Tiwari
Physics AI & simulation consultant

I find the real problem, build what is needed, and test whether it holds up.

Aditya Tiwari
Start with a Diagnostic See the study
Worked with
Endorsed by Dr. Manabendra SahariaIIT Delhi · ex-NASA Goddard Dr. Andrew TweedieCo-CTO, Quanscient Dr. Kostas LyrasMultiFluidX · KCL Dr. Eberhard BaischThermal Research Scientist
Find

Trace the fault before changing anything.

Paid client work2026

Get the simulation working.

Traced a moisture-to-vapour-pressure conversion error in MultiFluidX's OpenFOAM-to-ML pipeline, and kept their system.

Endorsed by Dr. Kostas Lyras

48reproducible OpenFOAM cases
1conversion error fixed
Phase 2paid follow-on
Held-out façade case Large E08: CFD temperature, POD-GP prediction and signed error, on a plane at z = 6 metres.
Follow-on surrogate study: CFD temperature, prediction and signed error for held-out façade case Large E08 at z = 6 m.

View full comparison · all fields ↗

Research studyPower electronics

Where is solver time worth spending?

Sensitivity analysis before sampling a 50-parameter design space.

1,574solver calls
50parameters screened
2solvers: OpenFOAM, CalculiX
Bar chart of effective input counts by engineering response at 80%, 90% and 95% total-sensitivity thresholds.
Different outputs depend on different inputs: 4–9 inputs account for 90% of total sensitivity per response in this study.

Enlarge sensitivity chart ↗

View held-out prediction checks
Held-out surrogate results by engineering response, comparing R-squared and rank correlation; performance varies across outputs.
Predictions checked against held-out solver results. Performance varies by output; these are solver-based research results, not physical validation.

Enlarge validation chart ↗

Build

Make the tool, with checks built in.

Paid contractMar–May 2026

Surrogates for a multiphysics platform

Solver-validated surrogate work for Quanscient.

Endorsed by Dr. Andrew Tweedie

Gaussian-process surrogate
Active learning: sample where uncertain
Pareto front across objectives

Schematics of the methods used

Research collaboration2025

Yamuna flood model

Built from IIT Delhi HydroSense's SAR inundation data.

Endorsed by Dr. Manabendra Saharia

2 daysfrom data to working hindcast
≈ 77%IoU in the primary river corridor
Perspective map of the Yamuna flood extent across Delhi, Noida and Faridabad on 3 September 2025.
Yamuna flood extent · 3 September 2025 · perspective view

Enlarge flood view ↗

View SAR comparison and validation
SAR data · HydroSense Lab, IIT Delhi
Model–SAR flood comparison within the river corridor, plus or minus 200 metres: IoU 0.774, precision 0.790 and recall 0.975.
Model–SAR comparison · river corridor (±200 m). IoU 0.774; precision 0.790; recall 0.975.

2D hindcast · presented to a national disaster-management adviser

Test

Measure where it holds up.

Independent researchOpen source

When can a physics-AI model be trusted?

A field surrogate, tested against OpenFOAM inside and outside its training range.

Code and evidence pack on GitHub

848solver-qualified cases
90.8%coverage at the 90% target, in range
23automated tests

Bounded benchmark · not a replacement for OpenFOAM or a production model

In range (100)Out of range (202)Routed to OpenFOAMDrag to move τ

Routing policy · click a gate

  1. 302 held-out cases
  2. Envelope check
  3. Uncertainty ≤ τ = 0.56
  4. Surrogate answers
Mean 90% coverage, answered–

Fig. 1 Error vs. predicted uncertainty for every held-out case. Out of range means 3, 10 or 12 fins (trained on 4–8) or pushed dimensions. Uncertainty bands widen 3.7×, but barely rank the error (ρ = 0.12).
Temperature fields for an in-range and two out-of-range cases: OpenFOAM, surrogate, absolute error and 90% half-width. Error is about 2 K in range and 12–14 K out of range.
Fig. 2OpenFOAM vs. surrogate, error and 90% half-width (K). In range: ~2 K error. Out of range: 12–14 K.
Consulting

Start with one stuck case.

Diagnostic

Start here

$500–1,500

One stuck case → written findings, a readout, next steps.

Start a diagnostic

Project

$8–25k · 2–6 weeks

A reliable baseline, a fix, or a validated study.

Discuss a project

Retainer

from $6k / month

Ongoing simulation reliability and physics-AI support.

Discuss a retainer

USD or EUR. Fee and date confirmed after a quick look at your materials. A Diagnostic doesn't include a full fix unless agreed in writing.

Endorsements

Aditya took up our offer of sharing our SAR-based inundation data and they built an advanced flood model for the Yamuna floodplain in 2 days.

We’re happy with the progress, and identifying the code bug and the input/setup issue was particularly useful. The additional checks you carried out on the façade implementation were also helpful.

Dr. Kostas LyrasScientific Adviser, MultiFluidX · Researcher, King's College London

Thanks to Aditya Tiwari, who has been developing this solver-validated approach as an independent practice, for his work on this.

Your experience is exactly what we need for this phase. Thank you again for your excellent presentation.

Dr. Eberhard BaischThermal Research Scientist

Send me the case that won't behave.

hello@adityatiwari.in
Include if you can
  • Case directory
  • Solver logs
  • Expected vs. actual
  • What it feeds into
GitHub LinkedIn Résumé Writing