AI scientist · Seismologist · UT Austin

Zhe Jia

I train machine learning models end to end, from data pipelines and pretraining to fine-tuning, reinforcement learning, and evaluation. I also build models for physical sensor data and study how large earthquakes rupture. I am a Research Assistant Professor at the Institute for Geophysics, The University of Texas at Austin.

Portrait of Zhe Jia
At the University of Texas Institute for Geophysics.
Ground motion at station IU.MAJO in Japan, 76° from the 2023 Mw 7.8 Turkey earthquakevertical · 40 samples/s · 80 minutes · 0.05–8 Hz
Language models

The full LLM lifecycle, built from scratch

Three open repositories cover every stage of a language model, from crawling the data to reinforcement learning on a 24-GPU cluster. Every number is measured, and every failure is written down with its evidence.

Stage 1 · Pretraining

nanoGPT-Seis

jiazhe868/nanogpt-seis
IsoFLOP profiles and the compute-optimal frontier from 25 muP runs.

A 113M-parameter GPT pretrained from scratch on earthquake science and general text, with a crawler, MinHash deduplication, a BPE tokenizer, a GQA and RoPE decoder, two-GPU DDP training, and a scaling-law sweep.

Parameters
113M
Training tokens
822.7M
Hardware
2× A30, 6.5 h
Optimal size
N ∝ C0.69
Stage 2 · Post-training

nanoGPT-Seis-Posttrain

jiazhe868/nanogpt-seis-posttrain
The 113M model against general Qwen models on held-out seismology questions.

SFT and GRPO-style RL with programmatic verifiers, then executed equations so that an interpreter does the arithmetic. Held-out accuracy on numerical seismology questions rose from 4.7% to 27.9%.

Held-out accuracy
4.7% → 27.9%
vs Qwen3.5-2B
+10.4 points
Teacher
Qwen3.8-27B on 2 GPUs
Pitfalls documented
15
Stage 3 · Scale-up

qwen-seis-posttrain

jiazhe868/qwen-seis-posttrain
Where the answers go after SFT and RL on two numerical test sets.

Full-parameter SFT and multi-task RL of Qwen3.5-9B with verl, Megatron, and vLLM on 8 nodes of 3 A100 GPUs, with frozen test sets, two judges, contamination checks, and an account of 27 failures.

D-Num-hard
62.0 → 70.1
GPQA Diamond
56.6 → 74.2
Tokens per answer
−62%
Cluster
24 × A100-40GB
The Sensoformer architecture and its two-stage training pipeline.
The Sensoformer architecture and its two-stage training pipeline.
Scientific machine learning

Sensoformer

A set transformer that reads seismograms from any set of stations and returns an earthquake's magnitude and focal mechanism in one forward pass. Physics-structured domain randomization lets it learn from 100,000 simulations and transfer to real data.

  • 19.7° median error on held-out real earthquakes, equal to the uncertainty of the catalog labels
  • 78% smaller gap between simulated and real embeddings after domain randomization
  • 10% of the real labels outperform training from scratch on all of them
  • +31.8% more reliable focal mechanisms across the Southern California catalog
A large earthquake as a few subevents, and how stations in different directions see them.
A large earthquake as a few subevents, and how stations in different directions see them.
Earthquake source inversion

SubMIT

Bayesian subevent inversion for large earthquakes, and the method behind my studies of the Fiji, Ridgecrest, South Sandwich Islands, Noto, and Calama earthquakes. One command goes from downloaded waveforms to subevent models with uncertainties.

  • Parallel MCMC over the time, location, depth, and duration of each subevent, with moment tensors solved exactly
  • Automatic choice of the number of subevents, with a check that the choice is stable
  • Teleseismic and regional waveforms, with exact timing for each subevent
About

Research and background

I am a Research Assistant Professor at the University of Texas Institute for Geophysics (UTIG), where I lead a group working on machine learning for physical systems and on earthquake physics. Before joining UT Austin in 2024, I was a Cecil H. and Ida M. Green Postdoctoral Fellow at Scripps Institution of Oceanography, UC San Diego. I received my PhD in geophysics from Caltech in 2022, and my BS and MS from the University of Science and Technology of China, where I studied in the School for the Gifted Young.

My current work is on language models and scientific machine learning: pretraining and post-training LLMs, sim-to-real learning for sensor arrays, and neural surrogates for physical simulations. My earthquake research uses Bayesian inversion, high-performance computing, and dynamic rupture simulation to work out how large earthquakes start, grow, and stop. My publications include 23 papers and preprints, two of them in Science, with more than 1,300 citations.

I teach Environmental and Earthquake Seismology and Frontiers in Earthquake Physics at UT Austin and advise PhD students and postdocs. Teaching and mentoring · Publications · CV

News

Recent

  • Released qwen-seis-posttrain: SFT and multi-task RL of Qwen3.5-9B on 24 A100 GPUs with verl and Megatron, with every result and failure documented. Write-up.

  • Released SubMIT v3, a fully automated subevent inversion pipeline that goes from downloaded waveforms to final figures. Code · About.

  • Released nanoGPT-Seis and nanoGPT-Seis-Posttrain: a 113M GPT pretrained from scratch on earthquake science, then post-trained with SFT and GRPO. Write-up.

  • Sensoformer code, weights, and training data are public on GitHub and Hugging Face.

  • Our study of the 2025 Mw 8.8 Kamchatka earthquake, led by PhD student Junpeng Li, is published in Geophysical Research Letters.

  • A 3D model of the Izu-Bonin-Mariana subduction zone that links slab dynamics to earthquakes, led by Yida Li, is published in Earth and Planetary Science Letters.

Earlier news
  • Sensoformer, a sim-to-real model for irregular sensor arrays, is on arXiv.

  • Our method for measuring rupture directivity of small earthquakes from energy envelopes is published in JGR: Solid Earth.

  • Our study of the 2024 Mw 7.4 Calama earthquake is published in Nature Communications and featured by Jackson School of Geosciences News.

  • Our study of how large deep earthquakes grow is published in AGU Advances and featured as an Eos Editors' Highlight.

  • Joined the University of Texas Institute for Geophysics as a Research Assistant Professor.

  • Our study of the 2023 Kahramanmaraş, Turkey, earthquake doublet is published in Science as the issue cover, with coverage in UC San Diego Today.