Core Vision: Build an Electrophysiology Foundation Model and deploy AI Scientist agents for autonomous neuroscience discovery — powered by EuroHPC exascale infrastructure.
A collaboration between SNU Connectome Lab (Transconnectome I / Diver) and Viktor Jirsa's group (EBRAINS / INS, Aix-Marseille University).
Electrophysiology data (EEG, MEG, intracranial recordings) is massive, heterogeneous, and underutilized. Current analysis pipelines are task-specific, require extensive manual tuning, and don't generalize across datasets, devices, or populations.
A large-scale pretrained model that learns universal neural representations from diverse electrophysiology data:
- Data: Multi-modal (EEG, MEG, ECoG), multi-site, multi-device recordings — thousands of hours
- Architecture: Transformer-based with spatiotemporal tokenization (cf. BrainOmni, LaBraM, MEG-GPT)
- Pretraining: Self-supervised learning on raw neural signals — requires massive GPU compute
- Downstream: Zero-shot/few-shot transfer to disorder classification, BCI, cognitive decoding, biomarker discovery
Autonomous research agents (inspired by Sakana AI's AI Scientist) that:
- Generate hypotheses from literature and data patterns
- Design and run experiments on the foundation model
- Analyze results, produce figures, write manuscripts
- Iterate through tree-search-based exploration of the hypothesis space
- Scale with EuroHPC compute to explore thousands of experimental configurations in parallel
- Foundation model pretraining on 1,000+ hours of neural data requires sustained GPU-months
- AI Scientist agents running parallel experiment trees multiply compute needs 10–100×
- Institutional clusters (even DGX Spark) are insufficient for population-scale pretraining + autonomous exploration
- Chief Science Officer, EBRAINS AISBL
- Director, Inserm Institut de Neurosciences des Systèmes (INS)
- Co-creator of The Virtual Brain (TVB)
- Leads Virtual Brain Twin Project (2024–2027, €10M EU Horizon Health)
- Leads EBRAINS 2.0 WP3: Digital Twins through Modelling and Simulation
- Role in this project: Electrophysiology data access (EBRAINS datasets), simulation framework integration, clinical validation pipeline
- Transconnectome I — Diver Project: Connectome-based deep learning for brain disorders
- Expertise: fMRI/dMRI/EEG analysis, contrastive learning, brain network modeling
- Infrastructure: DGX Spark (local development), EuroHPC (target production)
- Role in this project: Foundation model architecture, AI Scientist pipeline, training infrastructure
| Date | Milestone | Status |
|---|---|---|
| 2026 May | Benchmark Access 신청 (monthly rolling) | Next |
| 2026 May–Jul | Foundation model scaling tests on LUMI/Leonardo | Planned |
| 2026 Jun | AI Scientist prototype (local DGX Spark) | Planned |
| 2026 Sep 4 | Regular Access 마감 (10:00 CEST) | Target |
| 2027 Jan | Regular Access 결과 발표 | — |
| 2027 Jan–Dec | Production: Foundation model pretraining (1M+ GPU-hours) | — |
| 2027 H1 | AI Scientist autonomous experiment runs | — |
| 2027 H2 | Joint publication with Jirsa group | — |
| Access Mode | Cut-off | 결과 발표 | 기간 |
|---|---|---|---|
| Benchmark | 매월 1일 (rolling) | 제출 후 2–3주 | 3개월 |
| Development | 매월 1일 (rolling) | 제출 후 2–3주 | — |
| Regular | 2026-09-04 | 2027-01 | 12개월 |
| Extreme Scale | 2026-10-19 | 2027-03 | 12개월 |
Phase 1: Benchmark Access (2026 Q2)
├─ Foundation model: single-node → multi-node scaling test
├─ AI Scientist: agent loop profiling, GPU utilization
└─ Target system: LUMI-G or Leonardo (GPU partitions)
Phase 2: Regular Access — Scientific Track (2026 Sep submission)
├─ 1,000,000 GPU-hours for full pretraining + experiment runs
├─ Benchmark results as scalability evidence
└─ 12-month allocation
Phase 3 (Future): Extreme Scale Access
├─ Population-scale foundation model (10K+ subjects)
└─ Continuous AI Scientist operation
→ See GPU Hour Justification for compute breakdown.
eurohpc-tvb-digital-twin/
├── README.md # ← You are here
├── proposal/
│ └── gpu_hour_justification.md # Compute resource breakdown (1M GPUh)
├── benchmarks/
│ └── scaling_plan.md # Strong/weak scaling test design
├── gpu_hour_model/ # Resource estimation models
├── scripts/ # Automation & data collection
└── docs/
├── eurohpc_application.md # Application guidelines & templates
├── jirsa_collaboration_plan.md # Jirsa/EBRAINS collaboration strategy
└── projects/
└── diver/README.md # Diver project integration
- BrainOmni — Unified EEG+MEG foundation model
- LaBraM — Large Brain Model (2,500h EEG pretraining)
- MEG-GPT — Transformer-based MEG foundation model
- Sakana AI — The AI Scientist
- AI Scientist v2 — Agentic tree-search for autonomous research
신규 참여자를 위한 단계별 가이드:
- 이 README를 읽고 프로젝트의 두 핵심 축(Foundation Model + AI Scientist)을 이해합니다.
- 관련 논문 리뷰: 위 Resources의 BrainOmni, LaBraM, AI Scientist v2 논문을 읽습니다.
- Diver 프로젝트 파악: docs/projects/diver/에서 커넥톰 기반 파이프라인을 확인합니다.
- Jirsa 협업 계획: docs/jirsa_collaboration_plan.md에서 EBRAINS 연계 전략을 확인합니다.
- EuroHPC 신청 절차: docs/eurohpc_application.md에서 요구사항을 확인합니다.
- 벤치마크 설계: benchmarks/scaling_plan.md에서 스케일링 테스트를 확인합니다.
- GPU 시간 산정: proposal/gpu_hour_justification.md에서 자원 요구량을 확인합니다.
Maintained by SNU Connectome Lab — Seoul National University, Department of Psychology.