P·A

My research asks one question: can inexpensive, targeted interventions make neural climate surrogates both fast and physically faithful? At the Climate Analytics Lab within Scripps Institution of Oceanography, under Duncan Watson-Parris, I built FASST, a compact atmosphere emulator that runs ~18× faster than larger autoregressive models and stays stable over a 140-year ocean-coupled simulation. Fast surrogates like these tend to lose physical variability, such as ENSO, when trained on MSE alone. I study low-cost fixes such as stochastic fine-tuning and probing & steering of a model's internal representations, drawing on my work steering language models at inference time in the Belkin Lab. My long-term goal is to build efficient, trustworthy ML models for atmospheric and ocean science.

Curriculum Vitae GitHub LinkedIn Email
Portrait of Pratham Aggarwal
fig.1 · Pratham Aggarwal, San Diego

research

Making neural climate surrogates fast and physically faithful.

One question organizes my work: can inexpensive, targeted interventions make neural climate surrogates both fast and physically faithful? Fast emulators cut the cost of running large climate ensembles, but trained on standard losses they quietly drop the physics that matters. I study cheap, inference-time ways to put it back, without the cost of retraining.

2025–present Climate Analytics Lab, Scripps ML Researcher

FASST: a compact atmosphere emulator for ocean coupling

Built FASST, a compact, non-autoregressive atmosphere emulator (7.7M parameters) coupled to the physics-based POP2 ocean. Its atmospheric inference runs ~18× faster than CAMulator, a 751M-parameter autoregressive emulator, and the coupled model stays stable over a 140-year simulation without SST restoring. Trained on MSE alone it reproduces damped ENSO variability and a cold bias in the tropical Pacific. That gap between strong regression skill and physical fidelity now drives my work on inexpensive interventions to recover lost variability.

  • atmosphere emulator
  • POP2 coupling
  • ENSO
  • physical fidelity

paper (arXiv) ↗ NeurIPS 2026 · Tackling Climate Change with ML workshop

2024–2025 Climate Analytics Lab, Scripps HDSI Scholar

ClimateBench 2.0: benchmarking climate models fairly

With Willa Tobin, under Duncan Watson-Parris, contributed to ClimateBench 2.0, an architecture-independent framework for evaluating physics-based and machine-learning climate models on a common set of predictive-skill metrics. Benchmarked 14 CMIP6 models across MAE, RMSE, and CRPS, a hands-on lesson that no single headline number captures skill, and that a model topping one metric can fail on another. Built efficient, at-scale preprocessing for NetCDF, Zarr, GRIB, and GeoTIFF data streamed from Google Cloud; work recognized with a $4,500 research scholarship.

  • CMIP6
  • CRPS / RMSE
  • xarray / Zarr
  • GCP

evaluation platform ↗ code ↗

2026 Belkin Lab, UC San Diego ML Researcher

Inference-time steering: a transferable fix for surrogates

A surrogate with high R² can still get the physics wrong, and retraining it for every flaw is costly and often counterproductive. With Dr. Qingsong Wang, I extracted topic-specific refusal directions from LLaDA-8B's internal activations to steer its behavior at inference time, without retraining; in my capstone (DSC 180A/B) with Professor Weng I extended this to make refusals explainable. I now want to carry this idea to climate: probing and steering FASST's internal representations to recover ENSO variability lost during training, while preserving long-term coupled stability.

  • activation steering
  • LLaDA
  • probing
  • interpretability

projects & awards

Selected work, built to ship and to win.

1st place · HackML · NSF HDR ML Challenge

Coastal-flooding risk prediction

Winning entry at HackML (UCSD SMASH, NSF HDR ML Challenge). Because floods are rare, I evaluated on precision, recall, and F1 rather than accuracy; across 30+ documented experiments a CatBoost–XGBoost ensemble performed best, reaching a 0.94 F1 on the final benchmark.

CatBoost · XGBoost · feature engineering · Jan 2026

1st place · Hackfrontier

Homeless-services CV platform

A geospatial forecasting platform integrating 35+ transit, demographic, and geographic features (67% siting accuracy), paired with a real-time computer-vision system (Oxen.ai, EyePop.ai) tracking transient populations for data-driven service allocation.

scikit-learn · OpenCV · Jun 2025

Internship · NIOV Labs

Steering an autonomous AI agent

Used activation steering to enforce structured outputs in an autonomous AI agent, and built an NLP pipeline extracting goals, tasks, participants, and deadlines from meeting transcripts at 88% agreement with human labels, plus automated scheduling that cut manual effort 60% with 95% successful confirmations. Containerized with Docker, deployed on Kubernetes.

steering · NLP · Docker · Kubernetes · 2026

Research report

Black-hole seed growth simulation

Simulated black-hole growth under Eddington and super-Eddington accretion, comparing light- and heavy-seed trajectories over 12 weeks of guided research in computational astrophysics.

numerical modeling · Astropy · read the paper ↗

news

Recent.

contact

Let's talk research.

I'm happy to hear from prospective advisors, collaborators, and anyone working on neural climate surrogates, ML for atmospheric and ocean science, or inference-time model steering. The fastest way to reach me is email.

praggarwal@ucsd.edu github.com/pratham-aggr linkedin.com/in/pratham-agg

toolkit: PyTorch · JAX · xarray · NetCDF / Zarr / GRIB · climate emulators · diffusion models · scikit-learn · Python / C++ / Java · GCP · Docker · Kubernetes · Git