Each physics problem gets the tool that's cheapest for the question at hand, some are reduced physics models with no training data, some are learned surrogates. Each entry below states where it stops working.

A short, intense prepulse ionizes the gas and heats its electrons through two competing channels: above-threshold ionization, which is collisionless and retains residual drift energy from field ionization, and inverse bremsstrahlung, which is collisional and randomizes quiver energy in electron-ion collisions. Existing models treat one channel well and the other poorly, and there is a broad density range where the two contribute comparably. LAPHI reduces the sub-picosecond interaction at each transverse point to coupled ODEs: ADK/BSI field ionization feeds a Drude model, closed self-consistently through an effective temperature and a single energy-density equation. It is the first self-consistent treatment across the full underdense range for arbitrary ellipticity, and it yields closed-form scaling laws rather than only numbers.

Strength: sub-cycle resolved, ~0.4 s per parameter point, within 2–3% of PIC/MCC benchmarks, no training data.
Limitation: reduced model bounded to the underdense non-relativistic regime. Sets the initial state; not a self-consistent kinetic run.
Status: publication in preparation. The forward map doubles as an indirect diagnostic for prepulse-driven ionization and preplasma formation during a campaign.

Once the prepulse has deposited its energy, the heated plasma expands over hundreds of picoseconds to nanoseconds, and that expanded profile is what the main pulse actually arrives into. Running full hydrodynamics for every candidate configuration is not affordable inside a campaign. PHYRE instead uses a three-parameter self-similar family for the expansion, which classifies the dynamics into five distinct regimes ranging from quasineutral flow to Coulomb explosion, and gives scaling laws for how energy partitions between species. The output predicts plasma conditions during the laser rise time and guides target and pulse design directly.

Strength: closed-form scaling laws instead of full hydro runs; covers overdense expansion and underdense channel formation in one framework.
Limitation: self-similarity is an assumption, not a derivation from the initial conditions; accuracy degrades where the profile is far from the similarity family.
Status: pre-print posted, arXiv:2511.02028.

Direct PIC is impractical at intra-campaign scale: resolving the laser wavelength across thousand-wavelength plasma lengths makes systematic parameter scans prohibitive, and for fast-ignition geometries the full 3D requirement has been estimated at hundreds of millions of node-hours. The GNS reformulates PIC as a graph problem, macro-particles as nodes carrying position, momentum, charge and mass; field-mediated interactions within a connectivity radius as edges, and replaces the field solve and particle push with learned message passing. Crucially it learns the local time derivative of the plasma state rather than a global input-to-output map, so predictions can be interrogated against physical metrics at every step. Built-in spatial and permutation invariance means it transfers across geometries and particle counts without retraining.

Strength: ~1000× speedup with roughly 100× fewer training cases than MLP, Gaussian-process, or DeepONet surrogates; semi-transparent; differentiable rollout enables gradient-based optimization.
Limitation: in-distribution it performs well; out of distribution it drifts, with error accumulating over long rollouts. Needs a nontrivial multi-code data pipeline and slow training.
Status: 1D+3V testbed, training data from four independent PIC codes so the network learns physics rather than code artifacts. 

Laser-plasma accelerator performance drifts, and the variables driving the drift, effective a₀, plasma density, chirp, are precisely the ones no diagnostic measures directly. The usual response is black-box optimization, which recovers performance without explaining what moved. This model separates the two halves of the problem: a transition model captures how the hidden state evolves shot to shot, an emission model maps that state onto measured beam metrics, and an extended Kalman filter inverts the pair to recover the hidden state. The result is a causal diagnostic, degradation can be attributed to a physical source in real time, not merely corrected for.

Strength: recovers hidden state from sparse shots; makes drift explainable and attributable rather than only compensated.
Limitation: bounded by the emission model, the inference is only as good as the assumed state-to-observable map. Hard to build well.
Status: validated against 8 hours of continuous LPA-FEL operation at BELLA-HTU, LBNL, exceeding 100,000 shots.

A working facility accumulates far more knowledge than any single operator holds, shot logs, alignment procedures, past campaign results, code documentation, almost all of it unstructured. Retrieval-augmented generation over a facility's own corpus, kept on-premise, makes that knowledge answerable during a run. The larger goal is an agentic layer where an orchestrator, a knowledge-base agent, and an execution agent can run the surrogate and return a recommendation inside a feedback cycle.

Strength: runs on the facility's own corpus and stays on-premise; connects operators to institutional knowledge that is otherwise held informally.
Limitation: retrieval quality is corpus quality, and the bottleneck has consistently been data rather than modeling, experimental data arrives unstructured, and organizing it has taken longer than the modeling did.
Status: proposed; agentic platform integration is a Phase I deliverable of the Genesis 13B effort with Tau Systems.