In a few days we shall be rolling out pre-release versions of Lanyon to customers and testers. If you are in Chicago for the @APSphysics's Division of Plasma Physics meeting, please stop by our booth, talk to @JunoRavin and Tasman Powis, run some cool simulations, and grab some merch!

There is a narrative that fusion is essentially a solved problem as far as physics is concerned, and only engineering challenges, like heat exhaust, designing neutron-damage resistant materials, etc, remain. However, this is not true. In a sense, of course this is trivially untrue as engineering *is* physics. It turns out that on the contrary, the physics of fusion specially as we approach the burning plasma limit (that is, the plasma self-heating is enough to sustain fusion) is still unknown. One reason is that a plasma is an incredibly multi-scale system, and problems that you fix at one scale then uncover or excite problems at other scales, in an almost fractal-like manner. In this deep-dive by @JunoRavin we start on a series of explorations using @lanyon_ai to study the multi-scale nature of the fusion problem, and, hopefully, explore if it is possible to stabilize the plasma long enough, and reduce transport losses to low enough values, to actually get to a sustained burn.

This is a highly non-trivial calculation: the copper projectile is moving at a speed of about 800 m/s and impacts a plate. The system is nonlinear with complex constitutive relations that account for elastoplastic deformations and flows. All linear and nonlinear waves are included. Now imagine replacing the square projectile with a conical one, moving at 1000 or 2000 m/s, getting close to the sound-speed in copper. What would happen? Would the plate stop the projectile or would it pierce the plate? The key here is that the complex physics that describes this phenomena, and the corresponding numerics, can now all be constructed in a few minutes. It is truly incredible!

This problem also plagues scientific machine-learning (SciML), in which neural operators are used to solve or approximate PDEs. We discussed the role of reporting bias in our paper on SciML a couple of years ago. nature.com/articles/s42256-0…

I am very excited to welcome Tasman as the newest member of @lanyon_ai. He adds critical strength to Lanyon, allowing us to build our our training and inference infrastructure. Very exciting!

This seems to be the earliest known documentary film on nuclear fusion. Incredible progress has been made in the last 75 years, but this film shows that the key fusion concepts already were intensely studied in the 1950s: Z-pinches, mirrors, stellarators, ion-traps, etc. A lot of the early work was done at @PPPLab, which at one point was the center of the plasma and fusion universe. After all, PPPL is where the stellarator was invented, major fusion machines were built, the foundations of modern gyrokinetic theory were laid, and most of the alogrithms in computational plasma physics were invented. My own group did incredible work in computational physics at @PPPLab. This excellence and energy has now shifted out of PPPL, sadly. However, the fusion program remains strong, with the action moving to private fusion companies, all of which trace their to the earliest time documented in this film. A historical artifact worth a careful watch! youtube.com/watch?v=nxqhUljp…

A big welcome to our first Research Fellow, Aanya Bhandari! Aanya will work on building our model architectures, training pipelines, and AI infrastructure. Excited to have her on our team!

I am very excited to work with Chris and @getjonwithit on building a type system for all of the physical universe!

We are building Lanyon to enhance human creativity and not to replace it. Hence, our vision is human-centered. I deeply enjoy doing physics, mathematics and computing, and want to build a tool that enhances this enjoyment and does not destroy it. In this post Jonathan (@getjonwithit) summarizes our vision, framing it as an evolution of human languages (both natural and formal) to represent abstract ideas. He also describes how we think of science both as a process and a product. In a way, one can think of this as internal/external split: the process of learning and manipulating abstract ideas is internal, while the applications of these manipulations are external. Both are human-centered. However, for us these are not two separate activities, but part of the same whole. I should add that this split is similar to the zahir-batin (exoteric-esoteric) concepts that occurs in many mystical Traditions across the millennia. AI, when properly done, is merely a reification of this ancient and perennial idea into a novel and modern technological guise.

Being able to construct new code and run simulations, all in the space of a few minutes, is terribly addictive. Almost like every simulation is a dopamine hit and I keep wanting more!

The Rasail Ikhwan as-Safa (The Epistles of the Brethren of Purity) is a great medieval encyclopedia, likely compiled by secret Neoplatonic or crypto-Ismaili study circles. Almost a thousand years later we still do not know by who or why these epistles were composed. The Case of the Animals against Men in the Court of the King of the Jinn is one of the longest sustained narrative in the collection. In it, animals sue humans and the case evolves into the question: what makes humans superior to animals? Almost all reasons the human lawyer gives are shown to be inadequate, and the case of the animals appears very strong. However, the King of Jinns still rules in the favor of humans. I recommend reading the whole narrative to find out why. We need a modern retelling of this tale: "The Case of the Artificial Intelligence against Men in the Court of the King of the Jinn". Clearly, the human argument for superiority can't be based on intelligence, or the ability to reason, or our superiority in logic and mathematics, or things of that sort. So what remains? I suspect that the answers will be not that different to the ones the Brethren discovered a millennia ago. I believe exploring this question is critical as we approach a turning point in human history. If we don't do this we risk putting our ourselves in the same position as T.S Eliot was in when he wrote The Waste Land, stating "I will show you fear in a handful of dust."

Bill Dorland was pioneer in computational plasma physics, and in particular gyrokinetic simulations of turbulence in magnetized plasmas. He was a rare physicist, specially given his seniority and prominence: he hacked away at code and did complex derivations, all the while managing a large number of graduate students, teaching, and running complex academic programs. He would often crash our GPU hackathons, implementing some new algorithm he had thought up. Bill was an incredibly kind person and had deep impact on many people. I was fortunate to get to know Bill very closely when he was hired at PPPL (@PPPLab) to create the Computational Sciences Department (CSD). Bill and I worked together, co-founding the department and setting it up as a research entity within PPPL. Bill became a friend and mentor. He was very supportive of my and my team's work, and helped us navigate the DOE @ENERGY and the national lab complex. This is not an easy task! You can read @JunoRavin's personal reflections on @lanyon_ai blog.

How does a stellar wind (a wind of charged particles emitted by a star) interact with planetary magnetic fields? At the global level MHD describes the overall evolution, though more complex models are needed to get key physics correct. For example, a multifluid ten-moment model will include the effects of anisotropic pressure and also the self-consistent evolution of Hall currents. In many situations full kinetic (Vlasov-Maxwell) simulations may also be needed. For example, see our previous work in Geophysical Review Letters, 46, 11584-11596. In this calculation we used Lanyon @lanyon_ai to simulate the interaction of a stellar wind with a planetary diopole field using the MHD equations. The generation of the code and simulation took only a few minutes of complete. One needs far more physics and larger simulations to properly understand the details of Earth's magnetosphere, for example, but already this MHD simulation shows overall qualitative details of how fields are advected and night-side current-sheet forms. Of course, one would not want to live on this planet! The stellar wind will quickly blow away any atmosphere, making it uninhabitable.

The Ideal magnetohydrodynamics (ideal MHD) equations are central in plasma physics. Almost all modern fusion machine designs are based on ideal MHD theory, specially for macroscopic equilibrium and stability. It also serves as the basis of more complex models that contain additional physics. In this technical deep-dive I show how Lanyon @lanyon_ai can create a solver with a single prompt ("create a solver for the 2D ideal MHD equations"), and once the solver is created one can run a whole host of simulations. This can be addictive! I simulated the flow around blunt bodies, the interaction of the solar-wind with a planetary dipole field, the stabilization due to shear flow in Z-pinches, and even a magnetic nozzle. All in a few prompts and a few hours of compute. Very exciting!

A couple of years ago Nick McGreivy (@NMcGreivy) and I published a paper in Nature Machine Intelligence on (mis)use of machine learning for solving PDEs. The key conclusion of this paper was that a lot of ML PDE literature is highly misleading as ML researchers tend to use weak baselines for comparisons, do not tackle production problems, and fail to report when ML does not work (which is very often). I do not think much has improved since then. I have seen too many failed and unscalable approaches, and a blind rush to build data-driven models just for the sake of it. We also pointed out the serious issue of perverse incentives. Sadly, this has actually become worse since then, especially with federal funding agencies pushing "AI/ML" without actually evaluating if this is a useful thing to do. Worse, scientists at universities and DOE National Labs (for example at my previous workplace, @PPPLab) have simply gone along (as they must as otherwise they will lose funding) and proposed "surrogates" (a term I really dislike) for absolutely random things, "foundation models" when no such things are needed to advance the science, and other highly unimaginative approaches. Without removing these perverse external incentives things are unlikely to improve. In general, data-driven approaches, though useful for certain narrow applications, will not lead to breakthrough algorithms for new and extreme-scale problems as humanity tackles some of the hardest physics and engineering challenges. Following up on the work with McGreivy, with @getjonwithit and @JunoRavin we started down the path of building formally verified solvers, realizing that to tackle the most complex problems one needs to return to the fundamentals: automated discovery of algorithms, neurosymbolic derivations of new asymptotic models, all with formal certificates of correctness to guide the discovery process. This also extends to formally bounded neural solvers, which we showed in our BEACONS work (arxiv:2602.14853). All of this led to the formation of @lanyon_ai in our quest to build the needed R&D pipeline to tackle these challenges. I am glad that many new and novel approaches are being tried. We need them (and more!) as we approach the very limits of the problems our present tools can tackle. However, let us not forget the lessons we have learned! nature.com/articles/s42256-0…

Building Lanyon AI @lanyon_ai with @getjonwithit and @JunoRavin has been one of the most exciting and energizing new chapters in my life. Things came together surprisingly quickly as we fleshed out our technical vision, and put together a company to turn our vision into reality. We are deeply thankful to our VC and angel backers. Partnering with them we will build out our neurosymbolic agent, and help solve some of humanity's toughest challenges, from aerospace engineering, space and atmospheric propulsion, and nuclear energy. And that is just the start. We shall soon make Leibniz envious. Exciting times are ahead!