Research

Columbia · IRI · NASA GISS · Fulton Ring

Models thatearn their skill.

Research across forecasting, foundation-model probing, physics emulation, climate-risk data, and applied AI for health.

R-01Jan 2026 — Aug 2026 · Manuscript in preparation (first author)

Subseasonal-to-Seasonal Monsoon Forecasting

IRI for Climate and Society, Columbia
with Prof. Andrew Robertson

Collaborators · Andrew Robertson · Muhammad Azhar Ehsan · Mohammad Erfani · Emile Dara Esmaili · Haoxi

11+

Model variants

~47%

Leak share of skill

0.039 vs 0.012

Best RPSS vs. ELR baseline

U-NetFiLMRPSSMJO · SST/Niño3.4 · IOD

01
Built FiLM-conditioned U-Net models for Week 3–4 Indian monsoon rainfall forecasting, ablating conditioning across MJO, SST/Niño3.4 and IOD with FiLM vs. broadcast conditioning. The best model (IOD+SST+MJO, FiLM) reached RPSS 0.039 — roughly 3× the extended-logistic-regression baseline (0.012).
02
Discovered a data-alignment leak — conditioning indices matched to the target window instead of the issue date — and built a per-sample diagnostic exposing it (~47% of apparent skill; RPSS 0.047 → 0.025). Proposed an operationally valid fix: conditioning on issue-time predicted indices (e.g., IITM's predicted MISO).
03
Replaced bootstrapping with deterministic leave-3-years-out cross-validation and evaluated skill per grid cell with spatial skill maps, documenting negative results without overclaiming.

Fig. 1 — Where the skill went

0.047

Reported

0.025

Leak-free

Conditioning indices were aligned to the target window, not the forecast issue date.

A per-sample diagnostic showed roughly 47% of apparent RPSS came from information a forecaster wouldn't have on issue day. The fix: issue-time predicted indices — operationally valid, honestly lower.

R-02Jan 2026 — Present

Foundation-Model Latents — NASA Prithvi-WxC

Columbia Climate School
with Prof. Andrew Robertson

Paper

“Subseasonal monsoon dynamics over South Asia revealed by the Prithvi WxC foundation model”

Tackling Climate Change with ML (TCCML) workshop · NeurIPS 2026 · Sydney · Dec 11, 2026

Paper ↗

NeurIPS '26

Venue

Indices

LOYO

Eval

Prithvi-WxCLSTMTransformerLatent probing

01
Built a latent-extraction workflow for NASA's Prithvi-WxC weather-climate foundation model, pulling daily encoder/decoder representations over South Asia and the Indian Ocean.
02
Trained LSTM / Transformer models to reconstruct subseasonal variability indices (MISO, MJO, IOD) under leave-one-year-out evaluation — testing whether the model's learned latent space captures subseasonal modes.
03
Accepted at the TCCML workshop, NeurIPS 2026.

R-03Fall 2026 — Present

ML Climate-Model Emulator

NASA Goddard Institute for Space Studies
with Dr. Anastasia Romanou's group

PyTorch

Framework

GISS ModelE

Base physics model

Emulation

Approach

Climate-model emulationPyTorchPhysics-based modelsSurrogate modeling

01
Building a PyTorch ML emulator on top of the GISS physics-based climate model, learning to reproduce its outputs at a fraction of the computational cost.
02
Training and evaluating the emulator against GISS model simulations to check that it stays faithful to the underlying physics.

R-04Ongoing

Climate-Risk Model Reconstruction

Fulton Ring (climate-data startup)

Collaborators · Co-founders Rajan Desai and Jeremy

FEMA NRI

Model

County

Grain

Case study

Output

Data lineageValidationGeospatial

01
Validating an agent-built reconstruction of FEMA's National Risk Index Future Risk model, recomputed with newer data, against FEMA's methodology.
02
Checking data lineage and spot-checking county-level results to make preserved public climate-risk data usable and trustworthy.
03
Co-authoring a case study on how projected risk shifts.

R-05Ongoing

Policy-Agent — Behavioral Policy Simulation

Columbia Mailman School of Public Health
with Prof. Sen Pei's lab

Cities

Census-calibrated

Population

vs. observed

Eval

LLM agentsSynthetic populationsSimulationCalibration

01
LLM-agent simulations of how populations respond behaviorally to public-health policy across three U.S. cities.
02
Built census-calibrated synthetic populations and evaluated simulated responses against real-world outcome data.
03
Working on making the agents' output distributions match observed behavior.

R-06Ongoing

Socratic AI Tutor for Radiology

Columbia University Irving Medical Center · Dept. of Radiology

Vision + LLM

Modality

Socratic

Method

Clinician HITL

Eval

Multimodal LLMsMedical imagingVision-languageHuman evaluation

What do you notice about the costophrenic angles?

The right one looks… blunted?

Good. What could blunt it — and what would you check next?

Illustrative dialogue

01
A multimodal LLM + vision tutor that guides medical students toward findings on X-rays through Socratic questioning instead of giving answers.
02
Includes human-in-the-loop evaluation with radiology clinicians to validate diagnostic soundness.
03
Related work: a vision-language system for chest X-ray report generation.

Publications

In print & in prep.

“Subseasonal monsoon dynamics over South Asia revealed by the Prithvi WxC foundation model”

Tackling Climate Change with ML (TCCML) workshop · NeurIPS 2026 · Sydney · Dec 11, 2026

AcceptedPaper ↗

“S2S Indian monsoon forecasting with FiLM-conditioned U-Nets”

First author

In preparation