Numerical simulation is the substrate on which most deep-tech engineering claims are now built, and the live questions are not whether it works but how far machine-learned surrogates can replace it, how much of a simulated result survives contact with hardware, and whether quantum or physics-based processors will ever take over the hardest simulation workloads.
Summary
“Simulation” here means computing the behaviour of a physical or engineered system from a model rather than measuring it. In practice this covers three quite different things that the sources treat as one continuum. First, classical numerical solvers: finite-element and multi-physics codes that discretise a domain and integrate governing equations. A study of hexagonal boron nitride as a heat spreader for optically heated gold nanospheres used finite-element simulation to quantify how flake thickness, in-plane thermal conductivity and interfacial thermal conductance affect cooling, then checked it experimentally with wavefront-microscopy nanothermometry. A thermomagnetic generator study built a digital twin coupling magnetic, thermal, fluid-flow and electrical domains, argued that three-dimensional treatment is necessary, and validated it against the highest-power-density prototype using only known geometry and material parameters.
Second, surrogates: cheap learned or reduced models trained on simulation output. A Deep Operator Network was trained on physics-based predictions of process-induced deformation in AS4 carbon fibre/epoxy prepreg to predict deformation probabilistically across non-isothermal cure cycles. A two-compartment model of affinity biosensors combining simplified mass transport, Langmuir kinetics and mass conservation reproduced finite-element results with more than a hundredfold reduction in computational time, and yielded closed-form expressions for equilibration time as a function of the Damkohler number.
Third, simulation as the target workload for new hardware. A coalition preprint on physics-based ASICs names materials and molecular simulation alongside AI inference and sampling as the applications that justify relaxing statelessness, unidirectionality, determinism and synchronisation in chip design ref. Quantum simulation is the same argument in a different technology: nanographene pi-systems are proposed as a scalable bridge problem between early and large-scale fault-tolerant machines, with a tensor-network method for Trotter eigenvalue errors and an observed error-cancellation effect that cuts quantum phase estimation circuit depth by roughly an order of magnitude. A Schrodingerisation-based Hamiltonian simulation of time-domain Maxwell’s equations was run on IonQ hardware, recovering signed field directions and agreeing with analytical benchmarks in two dimensions, with three dimensions still on simulators.
The parameters that decide the field are therefore: how tightly a simulation is anchored to measurement, how far a surrogate can be trusted outside its training distribution, and how large the gap remains between simulated and measured performance for whatever the simulation is being used to justify. That last parameter is the one most often left unstated.
Evidence base
- A transient multi-physics digital twin of a thermomagnetic generator, coupling magnetic, thermal, fluid-flow and electrical domains in three dimensions, was validated against experimental data from the highest-power-density prototype using only known geometry and material parameters (4 May 2026).
- A reduced two-compartment biosensor model reproduced finite-element binding kinetics with more than a hundredfold reduction in computational time and yielded analytical expressions for equilibration time and required volume versus Damkohler number (5 May 2026).
- Trotter error cancellation for energy differences between low-lying nanographene eigenstates gave approximately an order of magnitude circuit depth reduction for quantum phase estimation, while worst-case, average-case and eigenvalue Trotter error estimates varied by orders of magnitude (4 May 2026).
- The first quantum-hardware implementation of a Hamiltonian simulation algorithm producing signed vector-field solutions to time-domain Maxwell’s equations agreed with analytical benchmarks in two dimensions on an IonQ QPU, with three-dimensional results obtained on simulators (5 May 2026).
- A coalition preprint including an ARIA programme director and four US academic groups named materials and molecular simulation as target applications for physics-based ASICs, framed against unsustainable energy consumption and CMOS scaling limits (14 July 2025) ref.
- Headline performance for a two-layer all-optical InP neural network, 89.5% best-case MNIST accuracy at 10 GS/s, 47 TMAC/s and under 20 pJ per operation at full system level, was obtained from simulation rather than measurement (13 January 2022).
- A safe active learning framework for Ga2O3 sensor reliability was first evaluated in simulation, then demonstrated on an automated high-temperature probe station with a Pt/Cr2O3:Mg/beta-Ga2O3 device (5 May 2026).
Open questions
- Do learned surrogates such as the composites DeepONet carry usable error bounds outside their training distribution of cure cycles, and can they be certified for qualification decisions rather than exploration only?
- Does the Schrodierisation-based Maxwell solver extend from two-dimensional benchmarks on IonQ hardware to three dimensions on hardware, and at what circuit depth and error rate?
- Which Trotter error metric should resource estimates use, given that worst-case, average-case and eigenvalue estimates differ by orders of magnitude for the same nanographene problem?
- Do physics-based ASICs produce any measured speed-up or energy advantage on a materials or molecular simulation benchmark, as opposed to an architectural argument?
Assessment drafted 2026-08-31 from up to 18 KB sources using the technology-scorecard framework; scores are a draft read pending review.