Fangxin Dong M.S. student · School of Chemistry · Sun Yat-sen University

I tend to follow questions across disciplines.

I began in computer science, studied information and computing science, and crossed into chemistry for graduate school. I now work on AI for science at the School of Chemistry, Sun Yat-sen University.

The subjects have changed several times. The habit has not: when I care about a question, I keep learning my way toward it.

current work · in progress

Predicting charge transport across very different materials

One framework tries to predict carrier mobility across five very different material families. The relevant physical responses and scattering mechanisms differ across them, so a strategy that works well for one family may not transfer cleanly to another.

So this is not a black-box end-to-end predictor. The model predicts physically meaningful intermediate quantities, and transport physics does the rest.

Crystal and direction features predict effective mass, deformation-potential constant and direction-matched elastic modulus. Physical deformation-potential and Fröhlich–Feynman–Osaka optical-phonon calculations are combined through Matthiessen’s relation to obtain direction-resolved calculated hole mobility.
Machine learning predicts intermediate physical responses; transparent physical models calculate mobility from them.

Because the model predicts quantities that physical models can use, there is somewhere to look when a prediction is off. The intermediate responses help localize where a mismatch enters the physical pipeline.

a case from the same project

When improvement stops propagating

Literature-derived semantic features reduced MOF deformation-potential error by 13.4% in a controlled response-level probe. Replaying only that improved response through the transport model shows why a better intermediate prediction need not translate into a comparable mobility change.

Improved MOF deformation-potential prediction enters a controlled transport replay and its signal collapses. Median downstream attenuation is 5.36 × 10⁻⁴. A shared logarithmic mobility axis places optical-phonon mobility at 0.0266 and deformation-potential mobility at 39.52 cm² V⁻¹ s⁻¹; all 36 MOF direction rows are OP-limited.

Building for continuity

One of the most important things AI has given me is not faster answers, but a way to leave traces.

fragments
traces
retrieval
connections
a network
fig. 2 · fragments, traces, retrieval, connections, a network. drawn small on purpose.

A research thought rarely arrives complete.

So the working environment is built to keep partial thoughts alive through project memory, searchable conversations, experiment records, agent handoffs, and structured workspaces. These make fragments of unfinished thinking recoverable.

What started as external memory gradually became something closer to a navigable network: a way to return to a question after interruption and continue from somewhere other than zero.

Trying things first

I started using LLMs seriously in the GPT-3.5 era, when very few people in my chemistry lab were using them as part of research.

My role since has been less about prescribing a workflow than about absorbing the trial-and-error cost: testing models on niche AI4S and computational-chemistry tasks, comparing where each behaves well, experimenting with multi-agent coordination, and helping labmates diagnose why a workflow feels bad or fails. Everyone in the lab studies a different topic. What transfers between us is not a shared prompt, but a way of choosing.

AI fluency is less about finding a universal prompt and more about learning how to choose, combine, question, and adapt tools to a particular problem.

public work

Small things I wanted to exist

Framework Miner

An evidence-first literature-mining tool for MOF and COF synthesis papers. It extracts structured synthesis information while keeping each field attached to the evidence passage that supports it. The point is not to replace expert reading. Acceleration should not make evidence disappear.

Questions I keep returning to

  • What should remain human when AI becomes deeply embedded in scientific work?
  • If AI gives us back an hour, who gets that hour?
  • What kinds of context and continuity should AI systems help preserve?
  • What obligations might emerge as increasingly capable AI systems become more persistent participants in human work and relationships?

Elsewhere: Outside research, I have spent time maintaining online communities, organizing grassroots mutual-aid work, writing public-facing science communication, and contributing to anti-harassment advocacy.

sometimes I draw: