# Zenithon — Relevant Notes Compilation > **What they do:** Frontier ML for extreme physics — building foundational models for physics-based simulation, with a stated focus on nuclear fusion and high-fidelity engineering simulation. Founded 2025, UK (Axbridge). Early stage / stealth. ~$136K raised via King's Entrepreneurship Lab and Unruly Capital. > > Website: https://zenithon.ai/ --- ## Why This Is Relevant to You You have done original research in this exact space — your published work on hybrid quantum physics-informed neural networks for CFD (Sedykh, Podapaka, Sagingalieva, **Pinto**, Pflitsch, Melnikov, 2024) demonstrated 21% accuracy improvement over classical PINNs for 3D Y-shaped mixer simulations. Zenithon is trying to build the commercial layer on top of this research frontier. The core thesis: physics simulations are massively expensive (CFD, fusion plasma modelling, molecular dynamics), and ML-based surrogates / foundational physics models can accelerate them 1,000–10,000x. Zenithon is betting this becomes a platform, not a consultancy. --- ## Your Relevant Notes, Organised by Theme --- ### 1. ML for Physics Simulation — The Technical Core These are the most directly relevant notes to what Zenithon is building. **[[Neural Network Compression of Simulations]]** Training a neural network to approximate a physics simulation's output. Speed gain: 1,000–10,000x. A CFD simulation that takes hours compresses into a NN that runs in milliseconds. The real problem: how do you generate training data efficiently? Adaptive sampling (Bayesian optimisation, Delaunay triangulation) dramatically reduces cost. Key note: this is rapidly becoming commodity infrastructure — not a differentiator on its own. **[[Surrogate Models]]** Lightweight approximation of an expensive simulation. Physics sim = lossless file. Surrogate = JPEG. Most published methods work for 5–15 input variables. Industrial processes have 28+. Delaunay triangulation breaks above ~12 dimensions — production systems use deep ensembles or Gaussian processes instead. Key investment question: *what method are you actually running in production, and at what dimensionality?* **[[Simulation-Based Optimization]]** SBO uses a simulation model as the objective function for an optimiser. Two families: derivative-free methods (Bayesian optimisation, evolutionary algorithms) and reinforcement learning. Important distinction: continuous process optimisation and discrete scheduling are fundamentally different problem classes — watch for companies conflating them under one "AI" brand. **[[Auto-Generated Physics Models]]** The holy grail: automatically building first-principles physics/chemistry/thermodynamics models from facility data without manual engineering. Currently *unsolved at production scale*. Every company claiming it should be able to show one specific model that was auto-generated and validated against plant data. If they can't, it's aspirational. This is the difference between a consultancy and a platform. **[[Digital Twins]]** Physics-based twins (first-principles, accurate but slow) vs data-driven twins (fast but fragile). The holy grail is hybrid. The hardest problem is not building one — it's building the second one faster than the first. See also the auto-generation problem. Key players: Siemens (Simcenter), ANSYS, NVIDIA Omniverse, AspenTech, Azure Digital Twins. --- ### 2. AI for Mathematical & Scientific Discovery **[[AI for Mathematical Discovery - The Next Frontier]]** ← *Highly relevant* Your note covers the full stack of what's needed for AI-driven physics solving: - **Physics-Informed Neural Networks (PINNs):** Embed PDEs directly into the loss function. Can be trained with minimal data using only PDE residuals and boundary conditions. Mesh-free. Handle inverse problems. Key limitation: struggle with sharp gradients, spectral bias causes slow convergence for high-frequency components. - **Navier-Stokes as the canonical example:** No analytical solution. Every aerospace, automotive, chemical, and pharma company spends millions on CFD because of this. AI for physics is the direct unlock. - **Tensor Networks:** From quantum many-body physics. Decompose high-dimensional tensors into polynomial complexity. Natural interface with quantum hardware. The bond dimension gives a tunable expressiveness knob. - **Hybrid Quantum PINNs for CFD:** Your own published research. 21% higher accuracy vs purely classical NNs in complex 3D geometries. Quantum layers add expressibility for nonlinear dynamics. - **The verification bottleneck:** LLMs hallucinate 2.5–15% of the time. In formal reasoning for physics, this is unacceptable. AlphaProof (Lean-based) solves this via formal verification. > *"Whoever cracks formal reasoning at scale with verification, while leveraging physics-informed architectures for sparse data regimes, unlocks something closer to a general-purpose scientific discovery engine."* --- ### 3. Nuclear Fusion — Zenithon's Stated Application **[[Fusion Gain Factor]]** Deep first-principles breakdown of Q (the fusion energy gain factor). Key table: | Type | Q Value | |------|---------| | Scientific breakeven | Q = 1 | | Plasma ignition | Q ≥ 5 | | Net electricity possible | Q ≥ 10–20 | | Engineering breakeven | Q ≈ 22+ | | Commercial breakeven | Q ≥ 20–25 | The ML angle: accurately simulating plasma physics to push Q higher is enormously compute-expensive. Zenithon's foundational physics models could in theory accelerate the design loop for plasma confinement. **[[Nuclear Fusion and Alchemy]]** Fusion economics are negative ROI on energy alone — you currently put in more than you get back. But Marathon Fusion identified a surprising byproduct: certain mercury isotopes in the reactor blanket decay into gold. One 1.5GW reactor could produce 3 tons of gold/year worth ~$320M vs ~$266M from electricity. The broader lesson: *the mix matters, not each individual part.* The revenue model for fusion may be a bundle, not a single product. **[[The Finance of Alchemy]]** Deep sourced version of the above. If fusion can make more money from alchemy than from electricity, it changes the commercial breakeven math entirely. Gold's value as an inflation hedge may also be structurally affected by additional supply. --- ### 4. Physics Fundamentals — Your Background Depth **[[Fluid Simulation MoC]]** Your notes on the foundational structure of fluid simulation — from quantum mechanics → molecular dynamics → kinetic theory → continuum fluid mechanics. The central theme: *reducing information at every layer of abstraction.* Simulating fluids at a quantum level for even small systems is practically impossible — hence the need for model order reduction. This is the exact problem Zenithon is working in. **[[Quantum Chemistry]]** Molecular simulation aims to find a compound's ground state. Today's supercomputers can simulate up to ~22 electrons / 22 orbitals by FCI (full configuration interaction) — the pentacene molecule. Anything larger requires approximations. Qubits operate according to the same laws as the molecules being simulated, giving quantum computers a natural edge for chemistry. **[[Quantum x Battery Simulations]]** Phasecraft, Hyundai × IonQ (lithium-air batteries), Daimler × IBM (lithium-sulfur batteries) — all using quantum computing to model the dipole moments, ground states, and molecular behaviour of battery materials with higher accuracy than classical computing can achieve. **[[Quantum x Climate Modelling]]** Climate models are systems of PDEs (General Circulation Models). Solving them requires strongly simplifying assumptions. Quantum computing could help solve PDEs to obtain more precise solutions in reasonable time. Same structure as the fusion problem: expensive simulation of complex physical systems. --- ### 5. Deep Tech Framework — How to Think About This Company **[[The Deep Tech Growth Cycle is different]]** ← *Your own framework* Tech-led vs problem-led innovation. Deep tech companies succeed by merging both. The three value opportunities: 1. Improve performance of existing worthy applications (optimising existing processes with ML/physics-inspired methods) 2. Explore unsolved problems beyond current capabilities 3. Frontier de-novo opportunities Zenithon sits at #2 and #3. The question is whether they can bridge to #1 fast enough to sustain themselves. Your note's warning is apt: *"individual ecosystems get caught up in their 'deep tech' bubbles, over-indexing on never-ending benchmarking battles… customers care most about meaningful value, reasonable timelines, and usable results."* **[[Foundational Models MOC]]** Foundation models as a paradigm: large pre-trained models adapted to many downstream tasks via fine-tuning or in-context learning. Zenithon's stated goal is a *foundational model for physics* — analogous to GPT-3 for language, but for governing equations. Still very much an open research problem. --- ## Key Tensions / Questions to Explore 1. **Product vs research:** Are they building a product or a research programme? At $136K raised and 2 employees, this is effectively a pre-product research bet. 2. **The auto-generation problem:** Building one physics surrogate model is a services business. Building the platform to auto-generate them is the product. Which are they actually doing? 3. **Fusion specificity vs general physics:** Fusion plasma is an extreme, specialised physics domain. ML methods trained on plasma dynamics may not transfer to other physics regimes. Are they domain-specific or genuinely foundational? 4. **Where does your work intersect?** Your hybrid quantum PINN research is a direct predecessor to what they're doing. The question is whether the quantum layer adds enough value at current hardware fidelity to justify the complexity vs classical PINNs. 5. **The verification gap:** As your AI for Mathematical Discovery note notes — LLMs hallucinate. In physics simulation for safety-critical applications (fusion reactors, aerospace), unverified outputs are a liability. How does Zenithon handle formal verification? --- ## Connected Notes - [[Neural Network Compression of Simulations]] - [[Surrogate Models]] - [[Simulation-Based Optimization]] - [[Auto-Generated Physics Models]] - [[Digital Twins]] - [[AI for Mathematical Discovery - The Next Frontier]] - [[Fusion Gain Factor]] - [[Nuclear Fusion and Alchemy]] - [[The Finance of Alchemy]] - [[Fluid Simulation MoC]] - [[Quantum Chemistry]] - [[Quantum x Battery Simulations]] - [[Quantum x Climate Modelling]] - [[The Deep Tech Growth Cycle is different]] - [[Foundational Models MOC]] - [[Industrial AI MOC]] --- *Compiled: 2026-05-22*