Energy Physics Pipeline

Live window into the Zer0pa lab. Energy Physics Pipeline is an active in-silico pipeline workstream, not a finished commercial service.

Boundary: Research infrastructure for in silico energy science: electrochemical conversion (batteries, green hydrogen electrolysis, fuel cells, solid oxide cells, photovoltaics, thermoelectrics) and fusion / plasma physics. Outputs are research artifacts. No regulatory certification claims. No clinical or human-subject use. Defence / weapons applications are out of scope under operator policy.

What This Is

Energy Physics Pipeline is an in-silico electrochemistry and fusion pipeline: CPU authority is Runpod-ready; enterprise H100 completion remains active.

This repo is the canonical workstream surface for Energy Physics Pipeline: a six-layer research pipeline spanning electrochemical conversion and fusion / plasma physics. It contains the PRD, source briefs, CPU-first execution artifacts, falsification tests, Runpod readiness evidence, and the H100 completion plan.

The current state is deliberately WIP-visible. Wave 4 made the CPU substrate ready to accept GPU backends through the same public endpoints. It did not complete the enterprise GPU-backed pipeline. The remaining authority metric is the H100 execution wave described in H100-ENTERPRISE-COMPLETION-PLAN.md.

Pipeline Mechanics

Within Energy Physics Pipeline, the two sub-verticals may share L6 design and the L4 response schema. Across Health, Materials, and Energy Physics Pipeline, substrate sharing remains disallowed during build; redundancy is part of the parallel-exploration strategy.

Key Metrics

Repo Identity

Readiness

Honest Blocker

No GPU-backed enterprise completion wave has passed yet. A service smoke test, shaped envelope, or single endpoint flip is not completion. The first serious H100 mandate is one electrochem GPU lane, one fusion or reasoner lane, live same-endpoint cutover, audit/KG provenance, and full falsification/regression.

What We Prove

• Wave 4 same-endpoint cutover is wired: public /v1/<sub>/<layer>/<op> routes through energy_physics_pipeline.l6.backend_resolver.resolve_and_dispatch.
• Accepted envelopes go through central accept_envelope / accept_envelope_and_dro enforcement with production falsifiers, audit JSONL/DuckDB, and KG writes.
• The CPU-first suite is strict enough to reject overclaiming: full check, Runpod same-endpoint tests, mandatory audit/KG tests, and parallel audit/KG safety tests are present.
• Source evidence is content-addressed or demoted: current authority sources report 39 verified entries, zero authority failures, and two non-authority skips.
• The H100 completion plan names the real remaining work: dataful GPU-backed lanes, CPU-vs-GPU regression, cross-model disagreement, TDA where applicable, and enterprise handoff.

What We Don't Claim

• We do not claim the enterprise Energy Physics Pipeline is complete.
• We do not claim a Runpod smoke test, single shaped envelope, or one endpoint flip is sufficient completion evidence.
• We do not claim regulatory certification, deployable product readiness, clinical/human-subject use, or defence/weapons applicability.
• We do not claim all GPU/HPC tools are integrated; MACE/fairchem/eSEN, GyroSwin/CGYRO, vLLM reasoner, OpenMC/GPAW/R2S remain H100 work.
• We do not claim blocked or conditional licenses are cleared without kg://license-grant/... evidence.
• We do not commit bulk datasets to git; dataful execution must use manifests, small fixtures, checksums, and private object storage where needed.

Verification Status

Proof Anchors

Repo Shape

Sub-verticals

Energy Physics Pipeline spans two physically distinct sub-verticals that share the six-layer scale hierarchy and L6 orchestration spine:

• Electrochemical - Butler-Volmer master equation; polarisation curve V(j) is the device-response token; buyer context includes battery digital twins, PEM catalyst screening, perovskite PV, and SA PGM strategy.
• Fusion / plasma - Grad-Shafranov equilibrium plus gyrokinetic Vlasov-Maxwell master equations; plasma equilibrium state vector is the device-response token; IMAS-MCP enables an LLM agentic interface.

The orchestrator resolved the PRD structure as one PRD with Part A for electrochemistry and Part B for fusion, sharing L6 and the L4 output schema inside Energy Physics Pipeline only.

Front Door Receipts

Agent Read Order

For full execution context, read:

1. MODUS-OPERANDI.md - role chain and parallel-exploration principle.
2. PRD.md - long-horizon product requirements and interface contracts.
3. RUNPOD-READINESS.md - Wave 4 CPU substrate readiness for Runpod migration.
4. H100-ENTERPRISE-COMPLETION-PLAN.md - H100 enterprise completion gates and hour budget.
5. source-briefs/00-research-agent-handover-note.md - prior research handover.
6. synthesis/01-fresh-eyes-on-energy-briefs.md - fresh-eyes synthesis substrate.

Provenance

• Initial commit: 2026-04-30.
• Research input: electrochemical M2S brief; fusion / plasma second-pass brief; research-agent handover note.
• Synthesis/orchestration: fresh-eyes synthesis, PRD, Handoff-to-Overnight-Executor.
• Overnight execution: CPU-first build, Wave 2, Wave 3, Wave 4 hardening.
• Rename: Clean pre-public rename to Energy Physics Pipeline (2026-05-03).
• Current next role: H100 Runpod enterprise completion wave.

Cross-workstream Principle

This workstream runs in parallel with Zer0pa/Health and Zer0pa/Materials. Each workstream is built end-to-end as an independent pipeline. No substrate is shared during build. Redundancy across workstreams is deliberate: surplus coding capacity buys diversity of architecture. Any convergence happens later as a separate merge step after parallel workstreams complete.

Within Energy Physics Pipeline, electrochemistry and fusion may share L6 design and the L4 DeviceResponseObject; that is intra-workstream sharing and explicitly permitted.

Executor Build State

Overnight CPU-first build delivered. See FINAL-REPORT.md, RUNPOD-READINESS.md, and HANDOFF-FROM-OVERNIGHT-EXECUTOR.md.

Repo layout:

energy_physics_pipeline/
  schemas/                - UniversalLayerEnvelope, DeviceResponseObject, Falsification, Source, Reasoner
  audit/                  - JSONL + DuckDB writer with mandatory boundary check
  kg/                     - JSONL + NetworkX KG store; GraphML export
  rest/                   - FastAPI endpoints for L1-L5 electrochem + fusion
  l6/                     - config, backend resolver, registry, production falsifiers, enforcement
  tda/                    - persistent-homology early warning
  cli/                    - health, registry, smoke, serve-rest, falsification-wave
  adapters/electrochem/   - L1-L5 CPU, manifest, pointer, and Runpod-ready paths
  adapters/fusion/        - L1-L5 CPU, parser, manifest, reasoning, and Runpod-ready paths
  adapters/shared/        - source log, license gate, reasoner curator, Runpod dispatch
  mcp_servers/            - FastMCP servers for energy tool-calling
fixtures/
tests/
sources_log/
docs/decisions/
scripts/
tools/

Quick Start

git clone https://github.com/Zer0pa/Energy-Physics-Pipeline
cd Energy-Physics-Pipeline
python3.13 -m venv .venv
.venv/bin/pip install -e '.[test,tda,mcp]'
.venv/bin/pip install pybamm pybop pypsa pvlib cantera pyscf netCDF4 freegs omas pyrokinetics qiskit mcp ripser persim
ENERGY_AUDIT_DIR=$(mktemp -d) ENERGY_KG_DIR=$(mktemp -d) bash scripts/full_check.sh
energy-physics --help

Runpod migration starts by setting ENERGY_RUNPOD_BASE_URL and flipping the target layer with ENERGY_L?_BACKEND=runpod_rest. The enterprise completion standard is in H100-ENTERPRISE-COMPLETION-PLAN.md.

What's Next

The CPU substrate is complete. The enterprise completion wave is GPU-backed scientific execution on H100 via Runpod.

First mandate

Minimum enterprise completion requires one serious electrochem GPU lane, one fusion or reasoner lane, live same-endpoint cutover, audit/KG provenance, full falsification/regression, and a committed handoff. A smoke test or single shaped envelope is not completion.

H100 work plan

Full multi-lane completion (multiple electrochem + fusion lanes, calibrated dataful campaigns): 600–1,500 H100-hours.

Model and data dependencies

No bulk datasets commit to git. All data artifacts need a SourceManifest, checksum, and audit/KG provenance before entering the pipeline.

Three decisions before provisioning

1. Electrochem model: MACE-MP-0 (faster, broad materials coverage) vs fairchem eSEN (heavier, battery-specific). MACE is faster to bootstrap the first lane.
2. Reasoner: vLLM-served 70B on the same H100 (needs the full 80 GB HBM) vs a smaller domain model. Decide before the fusion/reasoner pod spins up — it drives instance count.
3. Data pointers first: Resolve source manifests (DIII-D scenarios, OQMD slices) CPU-side before GPU time starts. See tools/runpod_cutover_checklist.py and tools/verify_sources.py.

Operating constraints (GPU wave)

• Same boundary, falsifier, license, and audit/KG gates as CPU. No gate waivers for GPU speed.
• No Class C/D/E licensed tool enters the product path without kg://license-grant/... evidence.
• No bulk datasets in git; use manifests, small fixtures, checksums, private Hugging Face where needed.
• Set a 2-hour idle timeout on each pod and a hard budget kill switch before starting.
• Runpod operating runbook and artifact checksums must be committed before claiming enterprise completion.

Full gate definition: H100-ENTERPRISE-COMPLETION-PLAN.md.