Our paper proposing the Text-Twin-Translation (T³) framework has been accepted to the SIGKDD 2026 AI4Science Track (CORE A*). The same work was a Spotlight Oral at the ICLR 2026 AI4Mat workshop, presented recently in Brazil.

What T³ does

T³ is our answer to the data scarcity that pervades ML for complex nanomaterial/device applications. The workflow:

1. Automated prompt optimization via TextGrad. A text-gradient method drives an LLM-based agentic pipeline that extracts structured knowledge graphs from an unstructured publication corpus, cheaply and at high throughput.
2. Device-topology-aware Digital Twin. A Graph Neural Network that bakes in device-topology physical constraints, trained as a Digital Twin to predict coupled material–device performance.
3. Validation on an OOD downstream task. We apply the framework to designing and screening FET sensor probes for detecting PFAS in water.

We built it so the same pipeline can be pointed at other data-scarce material/device problems.

Links

• Code: github.com/ruiding-uchicago/T3_FET_sensor
• KDD OpenReview: openreview.net/forum?id=7QfGX651NZ
• ICLR 2026 AI4Mat workshop: sites.google.com/view/ai4mat/schedule

Thanks to my collaborators and to Schmidt Sciences for supporting this work.