Activate the power of your data and analytics.
Industrial reality is time-series telemetry, spatial imagery, spectroscopy, and 3D volume — not internet text. Physical Foundation Models are pretrained on cross-customer industrial substrate across all four. Two universal backbones, per-vertical FM portfolios, physics-aware extrapolation via FNO / PINN / neural operator, sovereign per-customer fine-tuning above shared pretraining — the structural moat operators on the platform get and competitors don’t.
The signals that decide industrial outcomes were never in the pretraining corpus
A senior reservoir engineer’s response to a drilling anomaly, a metallurgist’s read of LIBS spectra, a process operator’s adjustment from NIR + DCS + batch context, a grid operator’s dispatch under a contingency — the highest-value training signal in physical-world domains is not on the open web. Asking a general LLM to operate a refinery, a mini-mill, a foundry, a fab, or a substation is asking it to improvise.
Physical FMs are pretrained on industrial substrate that lives inside operators. Two universal backbones — a time-series / 1D-signal transformer that generalizes the Chronos / TimesFM / Moirai / TimeGPT / MOMENT / Lag-Llama lineage to industrial telemetry, vibration, and spectroscopy (NIR / FTIR / Raman / mass-spec / NMR / LIBS / XRF / OES); and a spatial-imagery transformer with 3D-volume extensions for seismic, geological block models, and CAD. Where extrapolation beyond operational experience is required — new crude type, FEOC-clean substitution, green-steel process, novel reactor condition — physics-aware FNO / PINN / neural-operator backbones handle it.
Three foundations of industrial perception
Two backbones. Physics-aware extrapolation. Sovereign moat.
Industrial AI fails when it tries to be one giant model trained on text. It works when it’s two universal backbones plus per-vertical FM portfolios, with physics-aware extrapolation for cases beyond operational experience, on a cross-customer pretrain substrate that’s sovereign per customer at fine-tune time.
Two universal backbones + physics-aware extrapolation
Time-series / 1D family (Chronos / TimesFM / Moirai / TimeGPT / MOMENT / Lag-Llama lineage, extended to DCS, vibration, spectroscopy, wireline). Spatial-imagery family (ViT + 3D-volume + spectral attention for hyperspectral and seismic). Where the operator has never run before — new crude, green-steel process, novel reactor — FNO / PINN / neural-operator backbones extrapolate from physics.
Per-vertical FM portfolios
3–5 FMs per vertical, tuned to where decisions live. Mining: Geological + Process + Fleet. O&G: Drilling + Reservoir + Refining + Methane. Power: Grid + Substation + Wind-Solar + BESS. A&D: Manufacturing + Engineering + MRO + Program. Chemicals: Process + Reactor + Material Qualification + Carbon Attribution. Steel: Mill + Metallurgy + Caster + Scrap. Contract mfg: Geometric + Inspection + Quoting + Cell-controller.
Sovereign fine-tune above cross-customer pretrain
Pretraining substrate trained on cross-customer cross-fleet data; per-customer LoRA / adapter layers stay in customer enclave with formations, recipes, toolpaths sealed. The structural moat: operators on the platform get cross-customer transfer learning that operators not on the platform never can.
The FM in action
Two backbones, one state representation.
Time-series telemetry, spectroscopy, spatial imagery, and 3D volume fuse into per-vertical state representations that every other North product — RL policies, edge inference, the protocols above — composes on top of.
Time-series + 3D-volume + spectroscopy → one state representation
Two backbones fuse with structured tabular context (MES, ERP, batch records, EVM). The output is state — what every downstream product needs to act.
3–5 FMs per vertical, tuned to the decision
Specialized portfolios beat one giant model on industrial tasks. Each FM is sized to the regime where its predictions actually drive value — and the underwriting-model licensing surface adjacent to it.
Geological FM (drill-core + assay + Leapfrog block-model + structural geology); Process FM (QEMSCAN / MLA particle imagery + flotation video + PI + LIMS); Fleet FM (haul-truck dashcam + LiDAR + cross-OEM telemetry). The KoBold pattern at industry scale.
Richest substrate. O&G: Drilling FM (WITSML + MWD/LWD + drill-string vibration spectra), Reservoir FM (3D seismic + production logs + reservoir-sim), Refining FM (DCS + analyzer NIR/GC/NMR + crude assay), Methane FM (hyperspectral + emission inventory). Chem: Process / Reactor / Material Qualification / Carbon Attribution — NIR + FTIR + Raman + mass-spec is the distinctive spectroscopy substrate.
Power: Grid FM (EMS + LMP + weather + outage + topology), Substation FM (drone + thermal IR + acoustic), BESS FM (cell-level BMS + dispatch + cycle history) — the underwriting model for the BESS-as-asset class. Steel: Mill / Metallurgy (LIBS + XRF + OES) / Caster / Scrap.
A&D: Manufacturing + Engineering + MRO + Program FMs (CAD geometry + CMM / NDT + MES + EVM). Contract mfg: Geometric + Inspection + Quoting + Cell-controller. Geological / Reservoir / BESS / Manufacturing FMs license to streamers, royalty firms, project-finance lenders, insurers as underwriting models — the FM-as-IP surface adjacent to operator subscriptions.
Per-vertical perception. Sovereign by default.
Two backbones, physics-aware extrapolation, per-vertical FM portfolios, cross-customer pretrain + sovereign fine-tune — the perceptual substrate every other North product runs on.