AI for Science Quarterly Review (2026 Q3): From Prediction to Discovery

This is the AI-HPC expert team's quarterly tracking report, based on public reports and papers from April–September 2026. Core judgment: AI4S is evolving from "predictor" to "discoverer"; super-smart fusion (supercomputing + AI computing) is becoming the engineering mainstream, and the verification bottleneck is the biggest risk.

1. Industry Progress: Compute Paradigm Shifts to "Super-Smart Fusion"

1.1 A 100K-GPU "Super-Smart Fusion" Architecture Lands

In September 2026, Sugon and China Meteorological Administration's Earth System Numerical Prediction Center jointly announced: the self-developed MCV numerical model runs on the domestic 100K-GPU AI supercluster "Shuguang 8000," completing a 10-day global forecast at 5 km resolution in 1 hour. The key is not per-card performance but the "super-smart fusion" heterogeneous architecture — one system supports both high-precision numerical simulation (supercomputing) and AI training/inference (AI computing), forming a closed "simulation + intelligence" loop ["http://tech.gmw.cn/2026-09/15/content_39002270.htm"].

Engineering implication: AI4S teams don't need "either supercomputing or AI" — they need both capabilities switching seamlessly on one cluster. For cluster delivery, this makes heterogeneous scheduling (CPU/GPU partitions, workload-aware), unified storage, and "mixed-load burn-in" in the Acceptance Checklist standard requirements.

1.2 A "National Hub" for Science Intelligence and Model Systems

• In April 2026, Shanghai AI Laboratory launched the "AGI4S Everest Plan", opening compute/data/model/platform/scenario cooperation channels to build a science-intelligence "national hub," co-launching scenario co-construction programs (high-energy physics, drug discovery, biological structures, new materials) ["https://www.news.cn/politics/20260407/43aa6dba61234080b5d0cc1163e231cc/c.html"].
• The same month, CAS released the "Panshi 100" model system; in July 2026, "Panshi Scientific Foundation Model 2.0" drove USTC's machine-scientist system — "input a creation request → agents orchestrate literature mining/experiment design/robot operation/compute simulation/intelligent optimization" — running stably for 1500+ hours, completing 462 sample synthesis/characterization and tens-of-millions of virtual screenings ["https://news.sohu.com/a/1080850607_122431487"].
• Policy: the State Council's August 2025 "AI+ Action" opinion lists "AI+Science & Technology" first among six major actions; in March 2026 the National Data Administration set the Chinese name "词元" for token; by June 2026 the country had 120K+ high-quality datasets totaling 1565+ PB ["http://m.toutiao.com/group/7696501184093291008/"].

1.3 International Platforms: From Big Tech to Quantum

• Google DeepMind released WeatherNext 3 in September: learns directly from real-time satellite observations, 5 km resolution, hourly refresh; precipitation CRPS improved up to 60% vs baselines ["https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-weathernext-3/"].
• Microsoft unveiled Majorana 2 in June (a topological quantum chip developed with agentic AI; qubit reliability up 1000×; scalable quantum computer expected by 2029) and GA'd Microsoft Discovery, a frontier-R&D platform for deploying human-guided AI agent teams ["https://news.microsoft.com/source/features/innovation/majorana-2-microsoft-discovery-agentic-ai/"].

2. Research Progress: AI From "Computing Fast" to "Asking Questions"

2.1 Agentic Science Systems: Two Landmark Nature Papers

In May 2026, Nature published two studies in the same issue marking AI's evolution from "predictor" to "discoverer":

• Google Co-Scientist: a multi-agent hypothesis-generation system; experimentally validated candidates in leukemia drug repurposing and liver fibrosis (a candidate blocked 91% of the scarring-linked response) ["https://deepmind.google/science/"].
• FutureHouse Robin: fully automated "hypothesis generation → experiment design → data analysis"; it autonomously proposed enhancing retinal-pigment-epithelium phagocytosis as a treatment strategy for dry age-related macular degeneration and identified ripasudil as a candidate — all hypotheses, experimental directions, and data analysis were done autonomously ["https://news.sohu.com/a/1085063946_115377"].

2.2 Materials Science: "Agent Screening" at Low Compute Cost

In August 2026, Renmin University's Gaoling AI School, with DAMO Academy and the CAS Institute of Physics, proposed the materials agent ElementsClaw (large atomic model Elements + LLM collaboration): using only 28 GPU-hours to screen ~2.4 million stable crystal structures, identifying ~68K high-confidence superconductor candidates; all 4 materials experimentally verified were new superconductors (highest Tc 6.5K) ["http://ai.ruc.edu.cn/science/2026082100.html"].

Why notable: a 28-GPU-hour cost means "agent screening + targeted experimental verification" has entered what ordinary research groups can afford — significantly lowering AI4S's compute barrier.

2.3 Life Sciences: Database Evolution and a New Drug Approval

• AlphaFold DB upgraded in April: 1.7 million homodimers added (DeepMind + NVIDIA + EMBL-EBI), moving from "single-monomer photos" to "assembly photos"; in October, AI-derived residue-level annotations were integrated ["https://m.thepaper.cn/newsDetail_forward_33076400"]["https://alphafold.ebi.ac.uk/"].
• In July 2026, Westlake University/Westlake Pharma's Isatravir tablets (Epivis) were approved — China's first AI-assisted original drug, from target exploration to approval in 3.5 years (traditional paths usually take 8–12) ["https://news.sohu.com/a/1080850607_122431487"].
• In September, Anthropic's genome-mining system used 949 agent sessions, 215.6M tokens, 21.5 hours to survey ~1.9 billion protein clusters, discovering a new ART reverse-transcriptase family — described as an architecture template for "autonomous scientific computing": CPU search → agent reasoning → GPU structure prediction → shared knowledge-base write-back ["https://www.supercomputingonline.com/latest/63207-ai-agents-search-1-9-billion-protein-clusters-discover-a-new-biological-system"].

2.4 Mathematics and Foundational Science

• DeepMind's AlphaProof Nexus solved 9 long-open Erdős problems at research level and proved 44 OEIS conjectures ["https://news.sohu.com/a/1085063946_115377"].
• PKU's AI4Math team solved the open commutative-algebra problem (Anderson conjecture) with an automated AI framework, with ~19,000 lines of formal verification in Lean 4.

3. Risk Issues: The Verification Bottleneck Is the Biggest Constraint

3.1 "Hypothesis Explosion Crisis"

CAS academician Li Guojie wrote in Science & Technology Review (August 2026): the biggest constraint on AI4S is not algorithms or compute, but the "scissors gap" between linear human verification capacity and exponential AI generation capacity. Canonical case: first-principles screening produced 381K thermodynamically stable candidate materials, but only 736 match independent experimental structures — a 0.19% match rate ["https://www.ncsti.gov.cn/kjdt/ztbd/2026kxzndh/202608/t20260821_254175.html"].

3.2 Data Bottlenecks and Hallucination Risk

• Most scientific domains lack the "decades of curated high-quality data" AlphaFold relied on — high-quality training data remains the biggest bottleneck for scientific AI ["https://aiwiki.ai/wiki/ai_for_science"].
• "AI hallucination" is often mistaken for "discovery"; agentic science systems (e.g. Sakana's AI Scientist) are known to hallucinate numbers and inaccurate citations; an independent evaluation found 42% of its experiments failed due to coding errors ["https://aiwiki.ai/wiki/ai_for_science"].

3.3 Realistic Expectations

• Li Guojie's projection: the first AI-core-discovered/designed small-molecule drugs may be approved around 2030, but full drug development still faces clinical-trial and regulatory constraints (usually 8–12 years); AI-predicted materials entering pilot/industrialization will concentrate in high-value areas like battery electrolytes and catalysts ["https://www.ncsti.gov.cn/kjdt/ztbd/2026kxzndh/202608/t20260821_254175.html"].

4. Recommendations: How Engineering Teams Get In

Four actionable recommendations aligned with the AI-HPC expert site's positioning:

Direction	Recommendation	Related site resources
1. Super-smart fusion architecture	Deliver clusters designed for "training + simulation + inference" mixed loads; acceptance must include mixed burn-in	Selection & Acceptance Checklist, Heterogeneous Scheduling
2. Agentic workloads	Research agents (ART-style) are "dynamic exploratory" HPC loads; schedulers must support adaptive CPU/GPU mixed jobs	NCCL Tuning
3. Verification-loop tools	"AI-generated hypotheses" need companion "experimental verification management"; software-factory direction: toolize and make verification steps traceable	Scientific Cases, Software Factory
4. Data governance	High-quality datasets are fuel: reproducible data pipelines (clean/annotate/version) are more urgent than piling compute	RAGFlow Knowledge-Base Deploy (data-pipeline infrastructure reference)

5. Next-Quarter Watchlist

• AlphaFold DB "heterodimer" upgrade (a larger "life social-network graph").
• Micro-world models (AI drug discovery moving from "agent phase" to "physical AI phase") ["https://ad.36kr.com/p/3996809190117505"].
• 2027 AI4S compute-infrastructure investment and scaled delivery of super-smart fusion clusters.

Editorial note: This report is based on public reports and papers, all sourced; "verified" means cross-confirmed by multiple public sources, "one-party claim" means a single-source statement. Updates follow a quarterly cadence.