Paving the Way Toward a Scaling Law for Scientific Discovery | Haohuan Fu | ADIA Lab Symposium 2026

Paving the Way Toward a Scaling Law for Scientific Discovery | Haohuan Fu | ADIA Lab Symposium 2026

Title history
  • Paving the Way Toward a Scaling Law for Scientific Discovery | Haohuan Fu | ADIA Lab Symposium 2026 — observed on 20261104-0952 UTC (YouTube)

by Haohuan Fu · View the channel on PSI

Formal & Physical Sciences AI — News & Applications UYComputer scienceUYQArtificial intelligence (AI)
🎙 Haohuan Fu 👥 879 📅 November 4, 2026 ⏱ 40 min 👁 0 📄 expert opinion 🧭 2026-11-04
Available in: English (current) Français

Keywords

scaling lawsupercomputingHPCAIDensity Functional Theory

Summary

In this presentation, Professor Haohuan Fu discusses the potential for scaling scientific discovery, drawing inspiration from the scaling laws observed in large language models. He argues that by concentrating resources—compute, data, and talent—into unified AI and HPC systems, science could accelerate breakthroughs. Fu introduces the LineShine supercomputer in Shenzhen, which combines double-precision HPC with AI capabilities, featuring a unique CPU architecture with 304 cores and HBM memory. He highlights three key applications: first, the XLSDFT framework scales Kohn-Sham DFT to 100 million atoms, enabling simulation of a battery interface at experimentally relevant scales and matching XPS measurements; second, numerical-AI climate ensembles with 10,000 simulations improve prediction accuracy; third, generative AI models compress Earth observation satellite data for efficient storage and access. These examples illustrate how algorithms, data, models, systems, and validation co-scale. Fu concludes by emphasizing the need to bridge computational and physical worlds, proposing a scaling law for scientific discovery where AI connects simulation, data, and experiments.

158 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides valuable insights into the convergence of HPC and AI for scientific breakthroughs. Fu demonstrates real, measurable results: a 100-million-atom DFT simulation and a 11-million-atom battery interface, which are unprecedented and validated by experimental comparisons. The argumentation for a scaling law is persuasive, supported by analogies from LLMs and concrete examples where scaling compute and data leads to qualitative improvements. The emphasis on co-scaling algorithms, data, models, systems, and validation is a nuanced perspective that avoids over-simplification. However, the claim that ‘science can scale’ is somewhat aspirational, as the examples cover specific domains and may not generalize universally. The presentation is technically solid, but the speculative extrapolation to a universal scaling law remains an open hypothesis.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is evident: the DFT work was an ACM Gordon Bell Prize finalist, and results were compared with physical experiments. The technical details about the supercomputer architecture are precise, and performance metrics are cited (1.19 exaflops FP64, sustained performance above 100 petaflops). Sources are limited to the symposium page; no external references are provided, but the presentation references foundational works like the Top500 list and mentions the Deep Potential AI model. The title is highly appropriate, as the talk systematically builds a case for scaling laws in science. The lack of comments precludes analysis of public reception.

232 words

Title / Content Match

The title accurately reflects the talk's focus on developing a scaling law for scientific discovery using HPC-AI convergence.

Quality & Reliability

8/10

The talk presents detailed technical results from the LineShine supercomputer, including a Gordon Bell Prize finalist work on DFT simulations, with clear methodological explanations. However, some speculative statements about scaling laws for all scientific discovery and reliance on AI-generated figures introduce uncertainty.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

Contribution & Novelties

The talk presents original results from the LineShine supercomputer, particularly the XLSDFT framework achieving 100-million-atom DFT simulations, which is a significant leap in scale. It also introduces a vision for a scaling law for scientific discovery, drawing parallels with AI scaling, and emphasizes co-scaling across multiple dimensions. This is actionable for researchers planning to leverage HPC-AI convergence.

Pour aller plus loin :

  • Density functional theory — Foundational method for quantum simulations, key to understanding the DFT scaling work.
  • Gordon Bell Prize — Recognizes outstanding achievements in HPC; relevant as the DFT work was a finalist.
  • Neural scaling law — Discusses how model performance scales with compute, data, and parameters, central to the scaling law thesis.
  • Exascale computing — Context for supercomputers like LineShine operating at exascale performance.

127 words

Radar Profile

The profile shows high scores in quality and technical level, with moderate scores in information quantity and global reliability, indicating a technically deep presentation with credible results but limited breadth and some speculative aspects.

Reliability 8/10