How we got here

• 2016 to 2022: AptML. A GPU numerical framework covering linear algebra, FFTs and neural networks, compiled from one source for either GPU or CPU.
• 2023 to 2026: Unified Physics simulator. Built on AptML, rather than combining various existing solvers for different scales.
• 2024 to 2026: Unified Chem graph transformer. Built on the same AptML framework to predict molecular properties, protein function, binding sites and drug–target binding affinity. It is in training now (see Appendix B).
• Three patent applications filed: the multi-scale engine, its quantum specialisation, and neural-network optimisations used by our Unified Chem AI tool.

Who we are looking for

We are looking for design partners in academia, startups and industry.

Our collaboration

• You bring one real problem and the objectives you would judge a result by
• A researcher or designer of yours works alongside us
• We co-publish what we find; you keep your design IP
• Your feedback decides what the simulator's interface becomes

What the toolchain looks like

• A separate package for every computational framework
  • Quantum chemistry, for electronic structure
  • Molecular dynamics, for conformations
  • Device or nanoscale tools, for transport
  • FEM or CFD, for the bulk structure around it
• Spreadsheets and scripts between each pair

What the seams cost

• Provenance. Continuum material parameters were measured or fitted somewhere else; difficult to trace them to physics
• Consistency. Each tool approximates in its own way, and chaining them leaves you unable to state how accurate the overall answer is.
• Observability. Other simulators show you converged field values, not where the energy actually went
• Elapsed time. The stitching between the tools is slow: days of setup, conversion and checking per handoff, repeated every design iteration

Provenance

• Pre-generate material properties in the same simulator: an orbital or molecular run produces them ahead of the nanoscale or continuum simulation that consumes them
• Each property is stored with its authorship and the configuration it was generated under, so the next user can judge whether it suits their problem
• Slide 5 covers this in full.

Observability

• Because the representation is a spectrum, energy per mode, harmonics, spectral entropy and quantum correlation are values you read directly
• Readable at any point in the mesh and at any step of the run, not reconstructed afterwards from a converged solution
• Cost per observation stays constant as the system grows
• Slide 6 covers this in full.

One element's spectrum. Each meridian is a phase, each parallel a frequency, with the equator the DC component. Colour and size track magnitude; here most frequencies sit near the DC phase and a few share it exactly.

Read at any point, any step

• Energy per mode: where energy sits in the spectrum, not just how much
• Harmonics: the mode content, live, as it shifts
• Spectral entropy: how disordered a local spectrum has become
• Quantum correlation: two-point correlation between any pair of locations

What it means at each scale

• Orbital: watch bonds form in time rather than inspecting a converged snapshot
• Molecular: follow energy through vibrational modes during a conformational change
• Nanoscale: see which phonon and EM modes carry heat or loss in a device
• Continuum: mechanical and thermal spectra in place, not from a separate modal analysis

What you bring

• A researcher or designer. One named counterpart who will actually use the tool
• Your target problem. A system from your own work, not a benchmark chosen to flatter anyone
• Your objectives. What a result has to show for you to call it useful, agreed in writing before we start
• Reference data where you have it: experimental, published, or your own prior calculations

What you get

• Access to all four scales, and material properties you generate yourself from a molecular structure
• Co-publication of results, with your reference data credited
• Your design IP stays yours. We claim nothing over what you model
• Direct influence on the UI. The interface will be minimal at first, by design; partner feedback decides what it becomes
• Direct contact with the people writing the solver, not a support queue

The representation

• Per element: mode amplitudes and phases, evolved in time
• 29 mode families spanning molecular, quantum and continuum regimes that each mesh element can choose from
• A correlated state needs no new data structure: it is the same element, holding fractional rather than whole occupation numbers

Locality and cost

• Nearest-neighbour coupling; global invariants by tree reduction
• Local observables at O(K) per point, with no eigensolver and no reconstruction pass
• GPU-native throughout, on the AptML framework begun in 2016
• No eigensolver, unlike LAPW, which implements a muffin-tin split, a static basis partition solving an eigenproblem, with no continuum coupling and no energy flux across the boundary

The architecture

• Multi-scale graph attention, with attention heads grouped by range: near neighbours, medium range, long range
• One architecture therefore covers a hundred-atom ligand and a thousand-residue protein without reshaping the model
• A unified node representation carrying type, chemistry, geometry and context
• Three-stage curriculum: molecular pre-training, protein adaptation with molecular rehearsal, then unified multi-task training

Why it sits with the simulator

• Its predictions are hypotheses. The orbital and molecular scales can simulate the candidates they point to, rather than scoring them with a second statistical model
• Simulation results then return as training data that the published literature does not contain
• Neither half is capped by what has already been measured and published