Before agentic, there was physics-informed

July 1st, 2026

At least from my vantage point in the world, I've been seeing physics-informed AI/tech get *relatively* less fanfare than agentic tech in recent months.† Yet this is a shame, because I think a physics-informed lens can be an enormous value-add to virtually any computational field. For those unfamiliar, physics-informed describes exactly what it sounds like: using physical models (i.e., baking in physical constraints) to explain oftentimes seemingly unphysical phenomena. Applications range from AI digital twins for energy infrastructure (ThinkLabs AI) to drug discovery (Schrödinger, Inc.).

Motivation

Physics-informed methods are attractive because they introduce inductive bias — creating predictive power for unseen events. In this sense, they are a compromise between pure simulation and blind extrapolation, achieved by encoding a conservation principle or a differential equation directly into the loss function (as in physics-informed neural networks) or into the dynamics themselves (as in neural ODEs).

The applications are far-reaching and impactful. In chemistry, equivariant machine-learned interatomic potentials like NequIP and MACE have reached quantum-mechanical accuracy from as few as ~100 reference calculations, which is up to three orders of magnitude (!) less data than their predecessors (Nature Communications, 2022). In fluid dynamics and engineering, neural operators like the Fourier Neural Operator have been shown to solve turbulent Navier–Stokes roughly 440× faster than the pseudo-spectral solver it was benchmarked against (ICLR 2021). In fusion, physics-informed neural networks reconstruct plasma equilibria, and in drug discovery, Schrödinger's physics-based free-energy methods predict binding affinity to within ~1 kcal/mol of experiment.

But there is one field in particular — aging/longevity research — that is intriguing to me as a whole, and I think physics-informed models may greatly benefit the area. I'll first dive into that, then step back into a more generalist framework for applying physics-informed techniques. And by the way, for those keeping score, you can count this as my contribution to longevity escape velocity.

A Physics-informed Case Study: Aging

Background

The historically underfunded and overlooked field of aging/longevity has fortunately been gaining traction in recent years on the backs of amazing funders like the Amaranth Foundation, age1, and Healthspan Capital (and many, many more!) and brilliant founders/companies like Loyal, Altos Labs, Retro Biosciences, and countless others that deserve discussion. The field has exploded in recent months as the delivery technology and the therapeutic modality are finally maturing at the same time — resulting in many known unknowns (and even unknown unknowns!) coming to light. We are at an inflection point in the history of aging/longevity research.

The modality stack has gone vertical, going from small molecules to biologics to RNA medicines (RNAi, ASOs, base and RNA editing) and epigenetic reprogramming. And delivery (historically the bottleneck) is finally yielding under the pressure of GalNAc conjugation (for the liver) and brain shuttles to breach the blood-brain barrier (see Aerska). This past January, the FDA cleared the first-ever human trial of partial epigenetic reprogramming (Life Biosciences' ER-100, which expresses three of the four Yamanaka factors); in 2024, Wave put the first RNA-editing drug into humans; and in 2025, BEAM reported the first in vivo base-editing correction of a disease-causing mutation in people.

But I posit that physics-informed models and methods still have enormous upside potential in revolutionizing the aging field. At first glance, this seems unlikely: aging appears…hostile to brute-force learning, to say the least. The sampling is sparse and heavy-tailed — i.e., you cannot densely probe a human across eighty years, the intervention life cycle is often decades long, and aging clocks are hard to measure and unreliable. The field seems like a potential cesspool for overgeneralized models learning nonsense from skewed data.

But here is where the physics-informed part comes in. It is an inductive bias for out-of-distribution behavior, or in layman's terms, a way of forcing the model to stay on track when data is scarce. And aging, improbably, hands you a near-invariant. Approximately two centuries ago (1825), Gompertz noticed that mortality hazard climbs roughly exponentially with age. So plot the log of hazard against age, and you get a line — could this be a conserved law?

That invariant is the crux of the physics-informed lens on aging, first explored in a 2021 Gero paper (Pyrkov, Fedichev, et al.). They essentially treated the body as a dynamical system fluctuating around an equilibrium and asked: once a perturbation knocks you off baseline, how long until you relax back? They named the answer resilience. And it decays, monotonically, with age (as you might expect) — the autocorrelation time and the variance both swelling and, extrapolated, diverging together somewhere between 120 and 150 years. This is critical slowing down: the universal spectral fingerprint of a system whose leading eigenvalue is creeping toward zero, its basin flattening until it can no longer recover. The restoring force that pulls you back gets weaker and weaker; recovery takes longer and longer (the autocorrelation time blows up), and random fluctuations grow larger and larger (the variance blows up). Read literally but cautiously, it is a bifurcation argument for a hard ceiling on human lifespan. To put it simply: as you age, your body becomes less resilient to diseases and such, and there comes a threshold around 120 where it is unable to recover.

My outlook

I think every dimension of the field of aging/longevity has room for physics-informed model optimizations, but I'll outline some of my favorites below.

1. Aging clocks. Fit a low-dimensional dynamical system (like a neural ODE on Gompertz's law) to sparse longitudinal data (like continuous glucose, heart rate variability, etc.) and read out the recovery rate, or the leading eigenvalue of the linearized dynamics. Gero has already productized the front end, estimating recovery rate from phone and wearable streams against 100M+ longitudinal records, apparently credible enough that Pfizer is using the platform to hunt fibrosis targets. But let's go further. The holy grail, I think, would be to turn recovery rate into a validated primary endpoint, compressing the feedback loop by orders of magnitude.
2. Interventions. Let's start by thinking about partial reprogramming as an engineering control problem on the cell-state landscape.‡ The physics comes from a quantitative cousin of Waddington's landscape (e.g. Waddington-OT, which reconstructs differentiation trajectories from time-course single-cell data), which you then use to solve for the trajectory: the minimal dose and timing of Yamanaka factors that nudge an old cell toward a younger basin without tipping it into full pluripotency. Companies like General Control are building the multiplexed epigenetic editors you'd steer with. And if you can read the leading eigenvalue continuously (see #1), you can control it by titrating interventions in real time.
3. Biopreservation. In other words, surrogating the physics of rewarming a vitrified organ. The hard part of cryopreservation is warming the organ back up without cracking it or letting ice re-form, described by a coupled bioheat/thermoelasticity PDE that can't be predicted a priori (peak stresses move around from organ to organ). But a neural operator (Fourier/PINN) trained on organ geometry (e.g. from a CT) and nanoparticle distribution could optimize the rewarming schedule to keep stress under the fracture threshold everywhere at once. The dream is per-organ vitrification as a solved design problem, eventually scaling from a kidney to the brain.
4. Discovery. A cluster of aging-relevant targets has been historically underserved by structure-based modeling: intracellular aggregates (amyloid, TDP-43), glycation crosslinks like glucosepane that stiffen tissue, oxidized cardiolipin in failing mitochondria. These are folding/binding/aggregation problems, which are the perfect prey for machine-learned interatomic potentials (NequIP, MACE) and free-energy methods.

Generalizations (i.e., why you should care)

Aging is one of many young and interesting fields that may be exponentially amplified by the introduction of more physics-informed approaches and models. If you're asking yourself whether physics-informed methods may be useful in your own field, here is my homemade litmus test:

• Is there a law to lean on?
  • A conservation principle, a rate equation, a symmetry, a decay curve, an invariant? It may be (and probably is) non-obvious, like the Gompertz curve in aging.
• Is data the bottleneck?
  • The sweet spot is data that is scarce, expensive, noisy, or sampled far too coarsely in time.
• Will violating a constraint invalidate your answer?
  • If your law is physical, this condition is likely satisfied. But it is still worth keeping in mind, as hard-coding "soft" laws can sometimes lead models astray.
• Is it secretly an inverse problem?
  • Often, you don't want to solve the equation so much as run it backwards — recover the rate constant, the diffusion coefficient, the hidden driving field consistent with both your sparse data and your law. Pure simulation can't (you don't know the parameter), and pure ML overfits.

If one or more of these conditions are satisfied, you may be in the market for a physics-informed approach. But as the double pendulum teaches us, before running your model, always remember to check your starting conditions.

Thanks to Martin Borch Jensen, Jesse Poganik, Nathan Lum, and Soleil Wizman for insightful conversations.