Ira Shokar

AI for Science & Engineering, NVIDIALondon, UK

Probabilistic ML for weather, physics and engineering.

Inductive biases Latent rollouts Scoring rules Uncertainty Generalisation Extreme events Weather & climate Aerodynamics
Ira Shokar

Probabilistic emulation

Scoring-rule training gives calibrated ensembles from weather to engineering flows.

Stable long rollouts

Latent rollouts and spectral losses keep emulators on the attractor for 105 steps.

Generalisation to extremes

Fine-tuning across parameters and out-of-training dynamics, for regimes never seen in the data.

Neural representations

Learned representations of geometry, satellite pixels and dynamics.

About

I work at NVIDIA on probabilistic neural emulators of fluid dynamics, with a focus on weather and climate modelling and aerodynamic engineering. My work spans research and application of AI models for physical systems, from probabilistic weather forecasting and Earth observation to aerodynamic prediction from CAD, with Earth-2 and AIE.

I am particularly interested in physical inductive biases for neural architectures (latent rollout models), optimisation and loss design for probabilistic modelling and uncertainty quantification (scoring rules, spectral losses), generalisation, and modelling extreme or rare events. If you would like to collaborate on any of these topics, do reach out: ishokar at nvidia.com.

Recent highlights

Experience

AI for Science & Engineering, NVIDIA

AI for fluids: weather (medium-range forecasting, downscaling, sub-seasonal to seasonal), climate emulation, and industrial CFD and FEA, including design optimisation and uncertainty quantification. I build scalable distributed data-loading, training and inference pipelines (hundreds of GPUs) within PhysicsNeMo and Earth2Studio, and translate research prototypes into production-grade implementations.

PhD, Applied Mathematics, University of Cambridge

Key research areas:

  • CRPS and rollout stability: CRPS loss, latent transformers and spectral regularisation give stable rollouts over 105 steps that match regime statistics.
  • Latent rollouts: autoregressive rollout fully in latent space, and the encode/process/decode losses needed for stability.
  • Scenario conditioning: conditioning on PDE parameters via AdaLN enables zero-shot interpolation across unseen forcing regimes.
  • Out-of-distribution and extreme events: emulators can forecast regimes and extreme events absent from training, and I characterise when this succeeds or fails.

Education

  • PhD, Applied Mathematics, University of Cambridge, Oct 2021 to Jan 2025
  • MRes, Applied Mathematics, University of Cambridge, Oct 2020 to Sep 2021
  • BSc Theoretical Physics (First Class), UCL, Oct 2017 to Sep 2020

Honours and awards

Teaching and service

Recent Publications

Ground-Wing: A Parametric 3D Wing-in-Ground-Effect Dataset for Benchmarking Neural Surrogate Models

A. Cenzato, L. Karch, A. Luckow, Y. Pichon, F. Pasqua, A. Ghasemi, others

arXiv preprint · 2026

A parametric 3D wing-in-ground-effect dataset for benchmarking neural surrogate models of aerodynamics.

arXiv

CANTO: CAD-Native Transformer Operators for AI-Aided Engineering

D. Leibovici, N. B. Kovachki, D. Ahn, R. Ohana, I. J. S. Shokar, A. Ghasemi, others

arXiv preprint · 2026

Neural operators that predict aerodynamics directly from CAD geometry and support differentiable design optimisation.

arXiv
CANTO: CAD-Native Transformer Operators for AI-Aided Engineering

Tessera v2: Scaling pixel-wise earth foundation models

Z. Feng, S. Jaffer, I. Shokar, J. Knezevic, J. Ball, P. Sousa, M. Elvers, M. Lisaius, others

arXiv preprint · 2026

Scaling laws and Matryoshka distillation for pixel-wise Earth observation embeddings, evaluated on 29 downstream tasks.

arXiv
Tessera v2: Scaling pixel-wise earth foundation models

Recent updates

November 2026

Ground-Wing, a parametric 3D wing-in-ground-effect dataset for benchmarking neural surrogate models, is on arXiv.

October 2026

Talk at ENGYS UGM 2026 on faster marine design with GPU-accelerated HELYX and AI surrogates.

September 2026

CANTO, CAD-native transformer operators that predict aerodynamics directly from CAD, is on arXiv.