An elegant idea, easy to oversell

A physics-informed neural network (PINN) starts from a genuinely beautiful move. Take a partial differential equation — the rule a physical field must obey everywhere — and instead of discretizing it onto a mesh, guess that the solution is a small neural network. Feed the network coordinates, read off the field, and use automatic differentiation to compute the derivatives the equation needs. Then train the network to make the equation's residual vanish. A boundary-value problem in calculus becomes an unconstrained optimization problem in the network's weights. No grid, no timestep, no linear solve.

Stated that way, a PINN sounds like a general-purpose solver that also happens to be mesh-free and differentiable. That framing is where the overselling starts. The plasma-PINN campaign has spent several phases asking the question the enthusiastic papers tend to skip: at matched accuracy, does the PINN actually win? The answer has been remarkably consistent, and it deserved a teaching artifact of its own.

The explainer

What a PINN is — and is not is a self-contained, offline visual field guide — the PINN story told in seven acts, in diagrams and math notation:

  1. The setup — a PDE as a residual f := u_t + 𝒩[u], zero exactly when the field solves the equation.
  2. The trick — replace the solution with a tanh network u_θ and let autograd supply every derivative, exactly, at any point, with no mesh.
  3. The loss — compose the data/boundary term and the physics-residual term into one scalar, ℒ_tot = ℒ_u + ℒ_f, and minimize it.
  4. The promise — mesh-free, sparse-data-tolerant, and a natural home for inverse problems where an unknown coefficient becomes a trainable parameter.
  5. The reckoning — the head-to-head against a competent classical mesh: ~10⁴–10⁵× slower on the forward solve, fragile under noise on the scalar inverse, and — the twist — its one apparent win on the ill-posed field inverse evaporates once the classical baseline is given the same smoothness prior.
  6. The proof — the part that turns anecdote into theorem.
  7. The moral — the prior did the work, not the network.

It closes with a plain two-column ledger of what a PINN genuinely is and is not, each claim tagged with its grounding: the canonical formulation, the convergence theory, or the campaign's own measurements.

The part worth pausing on: overfitting is a theorem

The most important slide is the proof. It is tempting to read the campaign's benchmark losses as a matter of tuning — surely with the right architecture and enough training the PINN closes the gap. The convergence analysis of Doumèche, Biau & Boyer (On the convergence of PINNs, arXiv:2305.01240) says otherwise. For standard PINNs they prove non risk-consistency: there exist minimizing sequences whose empirical (training) risk goes to zero while the true risk goes to infinity, with the weights ‖θ‖₂ diverging.

The mechanism is visual, and the explainer draws it. The residual is only sampled at finitely many collocation points. A network can thread every data point with a near-zero derivative at those points — passing the PDE test — and swing wildly between them. In the pure-solver heat case it is starker still: the function that is identically zero satisfies the linear PDE and the zero boundary conditions perfectly, minimizing the residual while completely ignoring the initial state it was meant to evolve. And the pathology is generic — it holds for friction, advection, heat, Schrödinger, Maxwell, and Navier–Stokes operators.

The cure is exactly the campaign's recurring theme, now as a theorem: regularize. Penalizing ‖θ‖₂ (ridge / weight decay) bounds the network's smoothness and restores risk-consistency; a stronger Sobolev-type penalty is what buys genuine convergence to the physical solution. The prior is load-bearing at every level.

The through-line

Put the benchmark and the theory side by side and the same sentence falls out of both. In an ill-posed inverse problem, what tames the problem is the prior — not the function approximator. And once you have the right prior, the cheapest vehicle for it wins: a mesh with Tikhonov regularization, or a Gaussian process. The same lesson shows up independently in production fusion, where Landreman et al. (2026) optimize stellarator confinement with a data-informed prior and no neural surrogate anywhere in the loop.

None of this says PINNs are useless. The honest niche is feasibility and convenience — a mesh-free solver you can stand up in a few lines where no gridded solver is set up, and a field you can differentiate anywhere. What the equations promise as performance, a competent mesh keeps taking back.

If you want the full campaign — six progressive PDE phases and three matched-accuracy head-to-heads — the case study has the numbers and the interactive deck has the walkthrough. Start with the visual field guide; it is the fastest way to see what a PINN really is.