.jpeg)
Physical Intelligence · Research Programme
Teaching machines how
physical systems behave.
Pythology is researching vertically specialised models that learn from engineering geometry, simulation, testing and field data while remaining constrained by physical principles.
Core thesis
The most valuable industrial data is not on the public internet.
It sits inside engineering organisations as geometry, simulation runs, test measurements, process records, maintenance histories and failure reports.
The programme does not begin with a claim to replace CAD, CFD, FEA, certification or physical testing. It begins with a narrower question: can a learned model accelerate one defined engineering decision while exposing its uncertainty?
Competing hypotheses
Three possible paths.
One current conclusion.
The programme is structured to reject weak assumptions rather than convert them into product claims.
Aerospace propulsion is the best first wedge
Turbine durability and engine health combine expensive failure, mature simulation workflows, measurable outcomes and potentially defensible proprietary data.
Fusion is highest-upside but too early
Plasma, materials, magnets, neutron damage and reactor operations create immense modelling value, but access and validation are unusually difficult.
A general physics model must start vertical
Foundation-model ambition can explain the destination, but engineering buyers purchase speed, accuracy, fewer tests and reduced downtime.
Leading research direction
Aerospace turbine durability and engine health.
The initial research target should be one measurable workflow—not a complete engine model.
Peak stress and temperature
Estimate high-risk thermal fields or stress concentrations under defined geometry and operating conditions.
Fatigue and creep risk
Classify damage accumulation or component-life bands using materials, loads and validated history.
Remaining useful life
Model degradation trajectories from operational settings, sensors and maintenance outcomes.
Technical architecture
Geometry, physical fields, time and uncertainty.
A useful model must connect design intent to simulated behaviour, physical validation, manufacturing variation and field outcomes.
Geometry encoder
Meshes, point clouds, design parameters, voxels or signed-distance representations.
Physics and simulation encoder
CFD, FEA, thermal, modal, fatigue and related field outputs.
Materials and process representation
Composition, properties, tolerances, manufacturing conditions and degradation behaviour.
Operational time-series encoder
Telemetry, loading cycles, inspections, maintenance and failure records.
Constraint and uncertainty layers
Boundary conditions, conservation, dimensional consistency, extrapolation warnings and calibrated uncertainty.
Data moat
Architecture can be copied.
Validated industrial history cannot.
The defensible asset is the relationship between design, simulation, physical testing, manufacturing and real-world performance.
Design intent
Geometry, materials, boundary conditions and design constraints.
Simulation history
Trusted solver outputs across successful, failed and abandoned designs.
Physical validation
Test-rig data, strain, temperature, inspection and destructive evidence.
Field reality
Operating cycles, maintenance, cracks, repairs and component replacement.
Validation framework
Speed without validation is not engineering intelligence.
Every model should be assessed against trusted baselines and explicit operational value.
Trusted comparison
Error against high-fidelity simulation and physical measurement.
Workflow speed
Time saved against the actual baseline engineering process.
New conditions
Performance on unseen geometries, materials and operating regimes.
Honest uncertainty
Whether predicted uncertainty corresponds to observed model error.
Falsification criteria
What would cause us to reject the aerospace-first thesis?
A research programme must be able to produce a negative conclusion.
Technical rejection conditions
Unpredictable error on new geometry; poor transfer from public or synthetic data; uncalibrated extrapolation; or physical constraints that do not materially improve reliability.
Commercial rejection conditions
Insufficient data access; integration cost greater than operational value; incumbent tools already solving the workflow; or every deployment remaining entirely bespoke.
Programme status
A controlled path from hypothesis to platform.
The current public status is Stage 01.
Stage 01 · Hypothesis definition — current
Define candidate tasks, buyer, data requirements, metrics and rejection criteria.
Stage 02 · Public benchmark prototype
Test degradation, surrogate and uncertainty methods on open datasets.
Stage 03 · Engineering validation
Compare outputs against trusted simulations and measured evidence.
Stage 04 · Design partnership
Apply the system to one controlled industrial workflow.
Stage 05 · Vertical model
Deploy a validated task-specific physical-intelligence system.
Stage 06 · Platform expansion
Extend shared representations across adjacent components and domains.
Scientific foundations
Methods and public benchmarks.
These sources establish technical precedent and benchmark data. They do not imply partnership or endorsement.
Physics-informed neural networks
Forward and inverse problems involving nonlinear partial differential equations.
Open paper →Fourier Neural Operator
Learning operators for families of parametric partial differential equations.
Open paper →DeepONet
Learning nonlinear operators through operator approximation.
Open paper →MeshGraphNets
Graph-network simulation over mesh-based physical systems.
Open paper →Digital Twin paradigm
Aerospace digital twins connecting simulation and vehicle health information.
Open source →Turbofan degradation data
Public simulated engine degradation trajectories used in prognostics research.
Open dataset →Manufacturing benchmarks
Controlled additive-manufacturing measurements for model validation.
Open source →Engineering collaboration
One dataset. One workflow. One measurable result.
Pythology is interested in discussions with organisations holding validated simulation, test, manufacturing, inspection or operational data for a narrowly defined engineering problem.
Pythology Physical Intelligence is an exploratory scientific and engineering research programme. Research outputs are not certified engineering analyses and must not be used independently for safety-critical design, operation or maintenance decisions. All outputs require task-specific validation by qualified domain experts.