What does AI model in optics?

Research note · Sources and products checked 12 September 2026

I wanted to understand what people mean when they say AI can model materials. Predicting a refractive index? Replacing an electromagnetic solver? Asking a group of agents to design a device?

All three are happening. They solve different problems, and putting them under one label makes the progress difficult to judge.

This is a reading guide focused on optics and nanosystems. The research findings below belong to the cited authors. The connections and proposed experiments are my working interpretation; they are not results from testing these products.

What is being modelled?

There are several places to put a learned model. Starting with the desired output makes the distinction easier.

1. Atoms → configurations. A learned interatomic potential predicts energies and forces. A simulation uses them to relax a structure or follow its dynamics.
2. Material → optical response. A model predicts the complex dielectric function, refractive index or an optical response tensor from a crystal structure or measurements.
3. Device → light. Given geometry, materials, illumination and boundaries, a model computes fields, scattering or a spectrum.
4. Target → next candidate. An optimizer or agent chooses another geometry, composition, calculation or experiment.

The first three describe different physical mappings. The fourth controls a search through them. MACE, OptiMate, neural field operators and photonic-design agents provide concrete examples of these different roles. [11] [8] [15] [18]

For a local, linear, isotropic and nonmagnetic material, its relative permittivity satisfies εᵣ(ω) = [n(ω) + i k(ω)]² with the exp(−iωt) convention. Here n is the refractive index and k is the extinction coefficient. An anisotropic material needs a tensor description.

Even a perfect prediction of that material response does not give the transmission of a patterned device. Its geometry introduces resonances, interference and diffraction. Conversely, a Maxwell solver supplied with a dielectric function does not discover the electronic transitions that produced it.

Where the numerical methods fit

The method comparison has the longer guide. For this question, the useful distinctions are:

• Layer matrices and RCWA: planar films, periodic gratings and layered metasurfaces. Meent and torcwa connect RCWA to automatic differentiation, making geometry optimization and training-data generation accessible from Python. Fourier-order and layer-slicing convergence still matter. [4] [3]
• FDTD: evolve electromagnetic fields in time. Meep, Tidy3D and Lumerical support adjoint inverse-design workflows. Broadband excitation is useful, but very fine features and slowly decaying resonances can be expensive. [1] [21] [2]
• FEM: solve fields on a mesh that can follow curved geometry and material interfaces. It is useful for resonances and coupled optical, thermal or electrical models. JCMwave combines optical FEM workflows with probabilistic optimization; COMSOL provides trainable surrogate functions alongside its simulation models. [23] [22]
• Boundary and volume integral methods: describe scattering through surfaces or a discretized material volume. MNPBEM and pyGDM are relevant to nanoparticles, nanoantennas and near fields; pyGDM also includes evolutionary geometry optimization. [5] [6]
• Electronic-structure and excited-state methods: calculate the material response itself. For example, Yambo’s Bethe–Salpeter calculations include electron–hole interactions in optical absorption. These can provide reference data for a learned material model. [7]

The choice of solver can matter more than the choice of neural architecture. Training a network to imitate an unnecessarily expensive calculation is one way to accelerate it. Choosing a better numerical formulation is another.

A related distinction is easy to miss: differentiable does not mean learned. An adjoint method computes how an objective changes with many design variables. A neural surrogate learns an approximation from data or a training objective. Both can support inverse design.

Learning optical material properties

The shift that interests me most is from predicting a single property to predicting a spectrum and its directional dependence.

OptiMate, by Malte Grunert, Max Großmann and Erich Runge, predicts dielectric functions and refractive indices from crystal structures using graph-attention networks. The 2024 paper introduces a database of 9,915 calculated dielectric tensors for semiconductors and insulators. Those reference calculations use the independent-particle approximation. [8]

GNNOpt also maps crystal structures to optical spectra. Its reported dataset contains 944 materials, and Kramers–Kronig relations connect the complex dielectric response to derived optical quantities. [9]

TSENN, by Ting-Wei Hsu, Zhenyao Fang, Arun Bansil and Qimin Yan, takes the tensor structure seriously. The 2026 work predicts frequency-dependent optical tensors while respecting crystal symmetry, using a dataset of 1,432 bulk nonmagnetic semiconductors. Rotating a crystal must transform the predicted tensor consistently. Encoding that requirement is a useful physical constraint. [10]

These dataset sizes are not a leaderboard. The studies use different targets and evaluation conditions. Accurate reproduction of a calculated spectrum also does not recover physics absent from that calculation. Strong excitonic effects, defects, surfaces and processing history need suitable reference models or measurements.

Atomic foundation models occupy another part of the picture. MACE-MP-0 learns a potential-energy surface and can be adapted to particular systems. That is useful for generating configurations and studying dynamics; energies and forces alone do not specify an absorption spectrum. An original benchmark found systematic softening in several broadly trained potentials, including MACE-MP-0, on tests involving surfaces and defects. Generality remains something to measure. [11] [12]

MatterGen moves from scoring structures to generating crystal candidates under property constraints. My interpretation is that this could feed an optical-material discovery loop. A generated structure with a desirable scalar property is still some distance from a synthesizable material with the required loss, dispersion and stability. [13]

Learned solvers and inverse design

A spectrum surrogate learns a mapping from geometry to selected observables. A field surrogate predicts spatial fields. A neural operator learns a mapping between input and solution fields across a family of problems.

Augenstein, Repän and Rockstuhl demonstrated a modified Fourier neural operator for electromagnetic scattering and free-form 3D inverse design in 2023. The attraction is repeated evaluation: train on a supported family, then predict many related cases cheaply. The cost of generating the data and training the network belongs in the comparison. [15]

MaxwellNet, by Lim and Psaltis, uses a residual of Maxwell’s equations as a training loss to predict fields for a family of micro-lenses. This illustrates physics-informed learning. It does not make every resulting prediction a converged numerical solution: sampled residuals, boundaries and interfaces still need checking. [14]

For inverse design, a network can also generate candidate structures. GLOnet, by Jiang and Fan, combines a conditional generative network with physical evaluation and gradients for dielectric metasurface optimization. The 2019 work is a useful example of learning a design distribution. [16]

Generating several candidates makes sense because inverse design is non-unique. Different geometries can produce similar spectra. The useful question is whether the candidate meets the optical and fabrication requirements when recomputed—not whether its pixels resemble one training example.

There are related notes on lossless physics-constrained surrogates and diffusion-based metasurface design in the article summaries. Their original source records and reading limits are retained there.

Agents, swarms and real experiments

“Swarm” can mean several rather different things.

Population search evaluates many candidate designs and updates or mutates them. Particle swarm and evolutionary algorithms belong here. They do not require a language model; pyGDM’s evolutionary optimizer is a concrete nano-optical example. [6]

Multi-agent reinforcement learning trains coordinated decisions. Mahlau and colleagues’ 2025 preprint assigns material/air choices across a photonic design grid and evaluates them using FDTD. Their algorithms outperform the gradient baseline they test. That is a result for those tasks and that comparison, not a general verdict on adjoint optimization. [17]

LLM agents operate at the workflow level: write a script, change a geometry, run a solver, inspect a result and decide what to try next. Kharel and colleagues’ 2026 photonics preprint demonstrates this across several device and layout problems. One particularly useful detail is in Appendix S6: the initial multiphysics reproduction failed, an expert corrected the reference setup, and autonomous optimization proceeded from there. The setup remains part of the scientific work. [18]

Robotic experimentation closes the loop through actual material. Rainbow, described by Xu and colleagues in 2025, combines perovskite-nanocrystal synthesis, optical characterization and Bayesian optimization. Its objectives include emission energy, linewidth and a proxy photoluminescence quantum yield. That proxy should not be silently reported as an absolute quantum-yield measurement. [19]

These examples suggest a useful division of labour: physical calculations and measurements provide evidence; learned models make repeated predictions or guide sampling; agents coordinate the work. More agents do not automatically supply more independent evidence.

Companies and products

There are companies building tools at each of these levels. The table separates direct photonics products from molecular/atomistic platforms and public research software. It is a capability map, not a ranking or an account of hands-on product testing.

Nine examples, grouped by what they model. On a small screen, scroll the table sideways; it also supports keyboard focus and arrow keys.

For metasurface or photonic-device work, Flexcompute, JCMwave and PlanOpSim are particularly relevant specialists to investigate, alongside Lumerical and COMSOL. For molecular optical materials, Schrödinger’s organic-electronics workflows ask a closer question. Matlantis and Orb models are relevant when atomic structure and dynamics are the missing part of the model.

Before choosing a product, I would ask for one example matching the actual material, geometry and observable. Which solver generates the reference? What is learned? What domain was tested? Can the final design be checked independently? A speedup without those details is difficult to use.

Startup histories, funding and clients

The product question led to another one: who is building these tools, who backs them, and who might actually buy them? The profiles below bring together the public company records. Missing figures are left missing; a confident-looking estimate would not improve the comparison.

How this developed

This is a sequence of selected examples, not a claim that any one group invented the whole approach.

1. Numerical optimization became easier to compose with optical solvers. The documented Meep and Lumerical adjoint workflows illustrate optimization built on physical field calculations. [1] [2]
2. 2019: learned design generation. Jiang and Fan’s GLOnet combines a generative model with physical gradients for metasurfaces. [16]
3. 2022–2023: learning fields and solution operators. MaxwellNet and the Augenstein–Repän–Rockstuhl work represent different ways to use neural models for electromagnetic prediction. [14] [15]
4. 2024–2026: richer material response. OptiMate, GNNOpt and TSENN move toward spectral and symmetry-aware optical predictions from crystal structures. [8] [9] [10]
5. 2025–2026: more of the loop becomes automated. The cited reinforcement-learning, LLM-agent and robotic-lab studies automate different decisions around numerical or experimental feedback. [17] [18] [19]

The lines overlap. Neural models have not made the numerical methods disappear. They create new ways to reuse their results and decide where to spend the next calculation.

What I would test

For this playground, a useful first comparison would be a small family of dielectric gratings. Keep the material model, illumination and target diffraction order fixed, then compare a parameter sweep, direct gradients and a trained response surrogate.

The existing 1D RCWA experiment and 2D periodic diffraction experiment provide context for the physical question. This proposed comparison has not been implemented or benchmarked here.

I would measure:

• Total calculation and training cost, not only inference time.
• Error in the spectrum, phase or diffraction-order power that drives the decision.
• Behaviour on held-out geometry families and near narrow resonances.
• Whether an independently recomputed final design retains its predicted performance.
• Sensitivity to geometric tolerances and uncertainty in the material’s optical constants.

Then add an agent with the same solver budget and fixed validation criteria. Does it find a better parameterization or resolve a failed run? Does it simply spend more simulations? Either outcome would be useful to see.

For the material question, the corresponding experiment is to hold out a semiconductor family and compare predicted dielectric spectra with suitable reference calculations, followed by a separate comparison to measurements. The gap between those two comparisons is part of the result.

References and reading limits

This is a curated research article, not a systematic review. Some findings are based on abstracts or selected manuscript sections, as noted below. No hands-on commercial product comparison or neural benchmark was run. Company profiles use public company, investor, university, government, counterparty and transaction-counsel disclosures, alongside explicitly labelled company-profile observations. Revenue, private valuations and financing details are not supplied where they could not be verified. Current prices, licences, support quality and accuracy on a shared test case were not evaluated.

The most useful reading route is the Meep adjoint tutorial, OptiMate and TSENN for material response, the neural-operator paper for field prediction, the photonics-agent paper including its failed attempt, and Rainbow for experimental feedback.

1. Meep contributors. Adjoint Solver. Documentation; accessed 12 September 2026. Official documentation; gradient method and constraints.
2. Ansys. Photonic Inverse Design Overview – Python API. Documentation; accessed 12 September 2026. Official documentation; FDTD, adjoint shape and topology optimization.
3. TORCWA contributors. torcwa: GPU-accelerated RCWA with automatic differentiation. Software; accessed 12 September 2026. Official repository; PyTorch RCWA.
4. Yongha Kim et al. Meent: Differentiable Electromagnetic Simulator for Machine Learning. 2024 preprint / software; accessed 12 September 2026. Official repository; NumPy, JAX and PyTorch backends.
5. Ulrich Hohenester and Andreas Trügler. MNPBEM – A Matlab toolbox for the simulation of plasmonic nanoparticles. 2011 preprint. Author abstract; original scope and assumptions.
6. Peter R. Wiecha. pyGDM — A python toolkit for full-field electro-dynamical simulations and evolutionary optimization of nanostructures. 2018. Author abstract; Green dyadic method and evolutionary optimization.
7. Yambo contributors. Two particle excitations. Documentation; accessed 12 September 2026. Official tutorial; BSE optical absorption.
8. Malte Grunert, Max Großmann and Erich Runge. Deep learning of spectra: Predicting the dielectric function of semiconductors. Physical Review Materials 8, L122201, 2 December 2024. Publisher abstract; 9,915 calculated tensors and IPA limitation. Open preprint: Open manuscript
9. Hung et al. Universal Ensemble-Embedding Graph Neural Network for Direct Prediction of Optical Spectra from Crystal Structures. Advanced Materials, 2024. Publisher abstract and retrieved text; open preprint: Open manuscript
10. Ting-Wei Hsu, Zhenyao Fang, Arun Bansil and Qimin Yan. Accurate prediction of tensorial spectra using equivariant graph neural network. Nature Communications 17, 3330, 2 March 2026. Publisher-indexed abstract and introduction retrieved; bulk nonmagnetic semiconductor benchmark. Full access was intermittent.
11. Ilyes Batatia et al. A foundation model for atomistic materials chemistry. The Journal of Chemical Physics, 2025. Publisher text; learned potential-energy surface and fine-tuning.
12. Bowen Deng et al. Systematic softening in universal machine learning interatomic potentials. npj Computational Materials 11, 9, 2025. Original benchmark; errors in energies and forces away from training-like configurations.
13. Claudio Zeni et al. A generative model for inorganic materials design. Nature 639, 624–632, 2025. Publisher-indexed text; crystal generation, not a general optical-property guarantee.
14. Joowon Lim and Demetri Psaltis. MaxwellNet: Physics-driven deep neural network training based on Maxwell’s equations. APL Photonics 7, 011301, 2022; preprint 2021. Author abstract; residual-trained field prediction. Journal DOI: 10.1063/5.0071616.
15. Yannick Augenstein, Taavi Repän and Carsten Rockstuhl. Neural Operator-Based Surrogate Solver for Free-Form Electromagnetic Inverse Design. ACS Photonics 10, 1547–1557, 2023. Author abstract and institutional publication record; open preprint: Open manuscript
16. Jiaqi Jiang and Jonathan A. Fan. Global Optimization of Dielectric Metasurfaces Using a Physics-Driven Neural Network. Nano Letters, 2019. Publisher abstract and author manuscript; open version: Open manuscript
17. Yannik Mahlau, Maximilian Schier, Christoph Reinders, Frederik Schubert, Marco Bügling and Bodo Rosenhahn. Multi-Agent Reinforcement Learning for Inverse Design in Photonic Integrated Circuits. arXiv:2506.18627, 23 June 2025, v1. Full HTML methods and benchmark table inspected. Code: Code repository
18. Prashanta Kharel, Amin Khavasi, Xinzhong Chen and Tyler W. Hughes. Autonomous agentic design for photonics. arXiv:2606.00915, v1, May/June 2026. Full HTML, discussion and Appendix S6 inspected. Code: Code repository
19. Jinge Xu et al. Autonomous multi-robot synthesis and optimization of metal halide perovskite nanocrystals. Nature Communications 16, 7841, 22 August 2025. Full article and metric definitions inspected; proxy PLQY distinguished from absolute PLQY.
20. Flexcompute. FlexAgent MCP. Documentation checked 12 September 2026. Product documentation; available integrations, not an independent accuracy benchmark.
21. Flexcompute. Inverse design plugin. Documentation checked 12 September 2026. Documented adjoint/autograd optimization example.
22. COMSOL. Surrogate Models for Fast Multiphysics Simulation. Product documentation checked 12 September 2026. Surrogate types and integration; vendor capability statement.
23. JCMwave. JCMoptimizer. Product documentation checked 12 September 2026. Bayesian optimization, parameter reconstruction and uncertainty analysis.
24. PlanOpSim. Metaoptics Design Software. Product documentation checked 12 September 2026. MetaCell and MetaComponent scope, API and design workflow. Does not establish a specific trained neural model.
25. Schrödinger. Modeling for Organic Electronics. Product documentation checked 12 September 2026. Optoelectronic properties and molecular workflows. Individual predictions can use different physical/ML methods.
26. Schrödinger. MS Informatics. Product documentation checked 12 September 2026. Automated learning of material-property relationships.
27. Matlantis. Main Functions and Features. Product documentation checked 12 September 2026. Commercial atomistic simulation with a learned potential; not a general optical-spectrum guarantee.
28. Orbital Materials. Pretrained models for atomic simulations. Repository checked 12 September 2026. Public Orb force-field implementations; no general visible-spectrum capability inferred.
29. Microsoft Research. MatterGen: official implementation. Repository checked 12 September 2026. Open research software, property-conditioned crystal generation.
30. SPINS Photonics. About us. Team page checked 12 September 2026. Founder roles and public team roster. A roster is not a payroll count.
31. SPINS Photonics. HyperWave. Product page checked 12 September 2026. GPU FDTD, adjoint optimization, Python interface and product availability.
32. SPINS Photonics / LinkedIn. SPINS Photonics company profile. Snapshot retrieved 12 September 2026. Company-reported size band, private status and product-launch posts; relative post ages are not exact event dates.
33. SPINS Photonics. Inverse Design: Grating Coupler. Tutorial checked 12 September 2026. Public executable workflow; not an independent customer case study.
34. Logan Su et al. Nanophotonic Inverse Design with SPINS: Software Architecture and Practical Considerations. 2019 preprint. Academic software lineage; not the company's incorporation date or corporate funding history.
35. Luminescent AI. Photonic Chip Design for Datacenter Interconnect and AI Computing. Company page checked 12 September 2026. Founder, application areas and stated 2025/2026 startup-program participation.
36. Luminescent AI. Product. Product page checked 12 September 2026. Differentiable Julia FDTD and development caveats.
37. Luminescent AI / LinkedIn. Luminescent AI company profile. Snapshot retrieved 12 September 2026. Self-reported 2024 founding, New York location, size band, community-cloud and premium-design statements.
38. Imperial College London / David Silverman. AI spinout Polaron to accelerate design of advanced materials. 16 January 2024. University announcement of the spinout and research origins.
39. Polaron. Polaron raises $8m to build the ‘intelligence layer for materials science’. 3 February 2026. Company fundraising announcement; investor identities, team and reported industrial use.
40. UK Department for Science, Innovation and Technology. British start-up wins £1 million AI prize for breakthrough slashing materials development from years to days. 19 March 2025. Government award announcement. The prize is distinct from equity financing.
41. Polaron / LinkedIn. Polaron company profile. Snapshot retrieved 12 September 2026. Self-reported founding/location; inconsistent size band and associated-profile count retained as a limitation.
42. AVL. Battery Cell Design: Insights Through AI and Simulation. Partner page checked 12 September 2026. AVL confirms the collaboration and technical workflow; does not disclose sales or contract value.
43. CuspAI. AI Materials Foundry and company information. Company page checked 12 September 2026. Founders, private legal entity, headquarters, reported aggregate funding and named ecosystem partners.
44. Hoxton Ventures. Our investment leading a $30 million seed round in CuspAI. 2024 announcement; retrieved 12 September 2026. Lead investor's own statement of seed amount and participants.
45. Latham & Watkins. Latham Advises on CuspAI’s US$100 Million Series A. 7 October 2025. Deal counsel identifies the amount and co-leads; publication date differs from September announcement.
46. NEA / Lila Tretikov and colleagues. From Algorithms to Atoms, Part II: Doubling Down on CuspAI. 23 July 2026. Lead investor confirms $450m round, Kleiner Perkins co-lead, Bezos participation and MIRA/kUPS workflow. Investor perspective.
47. Zero Prime Ventures. Cusp AI Launches ‘AI Materials Foundry’ — and $450 Million Series B. 20 July 2026; syndicated 11 August 2026. Investor-authored announcement reports $2.6bn valuation. Indexed text retrieved; direct page returned 403. Not a public market cap or independently audited equity value.
48. Kemira. Kemira and CuspAI Forge Strategic Partnership to Pioneer AI-Driven Materials Innovation. 9 July 2025. Counterparty confirms partnership and initial PFAS-removal project.
49. CuspAI via Business Wire. CuspAI Launches ‘AI Materials Foundry’ a Global Network to Accelerate Breakthrough Discoveries. July 2026 announcement. Company-issued announcement: platform, customer/partner distinctions and named network. Membership does not establish paid contracts.
50. CuspAI / LinkedIn. CuspAI company profile. Snapshot retrieved 12 September 2026. Self-reported 51–200 band and 88 associated profiles; not audited employee headcount.
51. Alven. Altrove raises $10 million to turn AI-predicted materials into industrial reality. October 2025 announcement; article dated 3 November 2025. Lead investor: seed, cumulative fundraising, founders, industrial partnerships and stated commercialization targets.
52. Goodwin. Alven Capital invests in the €7.7 million Series Seed Financing of Altrove Tech. 4 November 2025. Deal counsel's euro amount differs from the investor's $10m headline. Do not count this as another financing round.
53. Altrove. Approach. Company page checked 12 September 2026. AI selection, synthesis and materials-development workflow.
54. Altrove. Altrove Demonstrates an End-to-End, High-Throughput Workflow for Next-Generation Inorganic Materials using Synopsys QuantumATK. Company technical article; checked 12 September 2026. Use of QuantumATK and laboratory workflow; using a vendor tool does not make the vendor a customer.
55. Altrove / LinkedIn. Altrove company profile. Snapshot retrieved 12 September 2026. Self-reported 2024 founding, Paris location, private status, 11–50 band and 35 associated profiles.