Renjie Li

Turning a scientific idea into an experiment often requires researchers to build, configure and repeatedly adapt the labo- ratory around it. This work demands resources and specialized expertise that can place ambitious experiments beyond reach. Here we introduce a framework for self-evolving laboratories (SEL), connecting computation, physical experiments and AI models through persistent cycles of observation and revision. GPT Astra uses camera observations and physical feedback to adapt robotic actions, digital twins and CAD designs. These changes are retained as reusable skills and infrastructure, allowing the laboratory to develop alongside its scientific tasks. In our fume-hood SEL prototype, GPT Astra identified the need for a beaker holder and storage stand and generated successive designs for on-demand 3D printing. Four holder generations and two stand generations culminated in parts integrated into the robotic workflow with operator assistance. A photoresist campaign proposed by Raven, an agentic AI platform for materials discovery, connects this platform to materials fabrication. Exposure to a 405 nm continuous-wave laser in the two-photon polymerization setup produced circular features that survived development; optical microscopy and reduced acrylate-band absorbance in ATR–FTIR support polymerization and qualitative double-bond conversion. The scientific workflow connects experimental outcomes and retained capabilities to the resources needed to establish and extend them. By making laboratory capabilities reusable and adaptable, SEL offer a path towards broader participation in experimental science, empowering researchers to build and extend the tools their ideas require. Project website: https://ai4ssi.github.io/self-evolving-superintelligence-lab/.

Turning a scientific idea into an experiment often requires researchers to build, configure and repeatedly adapt the laboratory around it. This work demands resources and specialized expertise that can place ambitious experiments beyond reach. Here we introduce a self-evolving laboratory framework that connects computation, physical experiments and AI models through persistent cycles of observation and revision. GPT Astra uses camera observations and physical feedback to adapt robotic actions, digital twins and CAD designs. These changes are retained as reusable skills and infrastructure, allowing the laboratory to develop alongside its scientific tasks. In a fume-hood prototype, GPT Astra identified the need for a beaker holder and storage stand and generated successive designs for on-demand 3D printing. Four holder generations and two stand generations culminated in parts integrated into the robotic workflow with operator assistance. A photoresist campaign proposed by Raven connects this platform to materials fabrication. Exposure to a 405 nm continuous-wave laser in the two-photon polymerization setup produced circular features that survived development; optical microscopy and reduced acrylate-band absorbance in ATR–FTIR support polymerization and qualitative double-bond conversion. An accompanying evaluation protocol relates the retention and transfer of capabilities to the time, cost and expertise needed to establish them. By making laboratory capabilities reusable and adaptable, this approach offers a path towards broader participation in experimental science, empowering researchers to build and extend the tools their ideas require. Project website: https://ai4ssi.github.io/self-evolving-superintelligence-lab/.

Large language model agents can connect scientific text, molecular generation, simulation, and laboratory planning, but their use in materials discovery is limited by weak reproducibility, hidden intermediate state, and scarce domain-specific quantum-chemical data. Here we introduce Raven, a state-aware multi-agent framework for materials discovery powered by QuantumChem-200K, a corpus of more than 214,000 organic molecules annotated with photoinitiator-relevant properties including two-photon absorption cross sections and singlet–triplet intersystem-crossing energies. Raven combines replayable state-graph orchestration with QuantaMind, a Qwen-2.5-32B model fine-tuned on QuantumChem-200K for molecular property prediction and screening. A frozen QuantaMind forward model verifies candidates generated by an inverse model, and GRPO post-training uses this verifier as a reward signal to align generation with target property constraints. In a two-photon-polymerization photoresin campaign, Raven generated, filtered, verified, and ranked candidate resins, connecting auditable agent reasoning to experimentally testable materials.

Graph neural networks (GNNs) are widely used for molecular property prediction, yet their effectiveness across application specific chemical targets, especially polymers, remains highly underexplored. In particular, for photo-sensitive polymers commonly used in photolithography, predictive accuracy depends on capturing subtle structure-property relationships involving nonlinear optical response, excited-state energetics, and physicochemical constraints. To this end, we present a systematic benchmark, called Polymer-Arena, of representative multi-task graph learning models for photo-sensitive polymer property prediction, covering classical message-passing, geometry-aware, and transformer-based architectures under a unified evaluation protocol. By controlling the dataset composition, molecular featurization, and hyperparameter tuning, we isolate architectural inductive biases that govern predictive accuracy, generalization, and stability. Our results reveal clear trade-offs between messagepassing depth, geometric encoding, and attention mechanisms, with no single architecture consistently dominating across tasks. Models incorporating directional or continuous-filter information excel on excited-state and nonlinear optical targets, while simpler message-passing architectures remain competitive on scalar physicochemical properties. These findings provide actionable guidance for model selection in datadriven polymer discovery and establish a reproducible baseline for future graph-based molecular learning. Dataset is available at https: //zenodo.org/records/18228299.

Large language models (LLMs) are rapidly pushing the limits of contemporary computing hardware. For example, training GPT‐3 has been estimated to consume around 1,300 MWh of electricity, and projections suggest future models may require city‐scale (gigawatt) power budgets. These demands motivate exploration of computing paradigms beyond conventional von Neumann architectures. This review surveys emerging photonic hardware optimized for next‐generation generative AI computing. We discuss integrated photonic neural network architectures (e.g. Mach–Zehnder interferometer meshes, lasers, wavelength‐multiplexed microring‐resonators) that perform ultrafast matrix operations. We also examine promising alternative neuromorphic devices and platforms, including 2D materials and hybrid spintronic–photonic synapses, which combine memory and processing. The integration of two‐dimensional materials (graphene, TMDCs) into silicon photonic platforms is reviewed for tunable modulators and on‐chip synaptic elements. Transformer‐based LLM architectures (self‐attention and feed‐forward layers) are analyzed in this context, introducing the mathematical operations associated with the transformers and identifying strategies and challenges for mapping dynamic matrix multiplications onto these novel photonic hardware systems. Overall, we broadly introduce state‐of‐the‐art photonic components, AI algorithms, and system integration methods, highlighting key advances and open issues in scaling such photonic systems to mega‐sized LLM models. We find that photonic computing systems could potentially surpass electronic processors by orders of magnitude in throughput and energy efficiency, but require breakthroughs in memory especially for long‐context windows and long token sequences and in storage of ultra‐large datasets, among others. This survey provides a comprehensive roadmap for AI hardware development, emphasizing the role of cutting‐edge photonic components and technologies in supporting future LLMs.

The efficiency of Bayesian optimization (BO) relies heavily on the choice of the Gaussian process (GP) kernel, which plays a central role in balancing exploration and exploitation under limited evaluation budgets. Traditional BO methods often rely on fixed or heuristic kernel selection strategies, which can result in slow convergence or suboptimal solutions when the chosen kernel is poorly suited to the underlying objective function. To address this limitation, we propose a freshly-baked Context-Aware Kernel Evolution (CAKE) to enhance BO with large language models (LLMs). Concretely, CAKE leverages LLMs as the crossover and mutation operators to adaptively generate and refine GP kernels based on the observed data throughout the optimization process. To maximize the power of CAKE, we further propose BIC-Acquisition Kernel Ranking (BAKER) to select the most effective kernel through balancing the model fit measured by the Bayesian information criterion (BIC) with the expected improvement at each iteration of BO. Extensive experiments demonstrate that our fresh CAKE-based BO method consistently outperforms established baselines across a range of real-world tasks, including hyperparameter optimization, controller tuning, and photonic chip design. Our code is publicly available at https://github.com/richardcsuwandi/cake.

In this work, we unveil a novel, to the best of our knowledge, AI-based design method (AIDN1) specifically developed for photonic crystal resonator designs, capable of handling complex designs with over 10 degrees of freedom (DoFs) and considering practical fabrication uncertainties to minimize the common simulation-to-reality (sim2real) gap. Especially, we introduce an ultrashort (Q-factor) of 2 × 10⁷ and maintains a theoretical Q-factor above 10⁵ even under high fabrication variations. Importantly, we emphasize that AIDN1 is generalizable and our work serves as a solid foundation for future laser fabrication endeavors beyond the realm of ultrashort 1D photonic crystal (PhC) resonators.

In the dynamic landscape of Artificial Intelligence (AI), two notable phenomena are becoming predominant: the exponential growth of large AI model sizes and the explosion of massive amount of data. Meanwhile, scientific research such as quantum computing and protein synthesis increasingly demand higher computing capacities. As the Moore's Law approaches its terminus, there is an urgent need for alternative computing paradigms that satisfy this growing computing demand and break through the barrier of the von Neumann model. Neuromorphic computing, inspired by the mechanism and functionality of human brains, uses physical artificial neurons to do computations and is drawing widespread attention. This review studies the expansion of optoelectronic devices on photonic integration platforms that has led to significant growth in photonic computing, where photonic integrated circuits (PICs) have enabled ultrafast artificial neural networks (ANN) with sub‐nanosecond latencies, low heat dissipation, and high parallelism. In particular, various technologies and devices employed in neuromorphic photonic AI accelerators, spanning from traditional optics to PCSEL lasers are examined. Lastly, it is recognized that existing neuromorphic technologies encounter obstacles in meeting the peta‐level computing speed and energy efficiency threshold, and potential approaches in new devices, fabrication, materials, and integration to drive innovation are also explored. As the current challenges and barriers in cost, scalability, footprint, and computing capacity are resolved one‐by‐one, photonic neuromorphic systems are bound to co‐exist with, if not replace, conventional electronic computers and transform the landscape of AI and scientific computing in the foreseeable future.

Optical resonant cavities with high quality factor (Q-factor) are widely used in science and technology for their capabilities of strong confinement of light and enhanced light–matter interaction. The 2D photonic crystal structure with bound states in the continuum (BICs) is a novel concept for resonators with ultra-compact device size, which can be used to generate surface emitting vortex beams based on symmetry-protected BICs at the Γ point. Here, to the best of our knowledge, we demonstrate the first photonic crystal surface emitter with a vortex beam by using BICs monolithically grown on CMOS-compatible silicon substrate. The fabricated quantum-dot BICs-based surface emitter operates at 1.3 µm under room temperature (RT) with a low continuous wave (CW) optically pumped condition. We also reveal the BIC's amplified spontaneous emission with the property of a polarization vortex beam, which is promising to provide a novel degree of freedom in classical and quantum realms.

Photonics inverse design relies on human experts to search for a design topology that satisfies certain optical specifications with their experience and intuitions, which is relatively labor‐intensive, slow, and sub‐optimal. Machine learning has emerged as a powerful tool to automate this inverse design process. However, supervised or semi‐supervised deep learning is unsuitable for this task due to: (1) a severe shortage of available training data due to the high computational complexity of physics‐based simulations along with a lack of open‐source datasets and/or the need for a pre‐trained neural network model; (2) the issue of one‐to‐many mapping or non‐unique solutions; and (3) the inability to perform optimization of the photonic structure beyond inverse designing. Reinforcement Learning (RL) has the potential to overcome the above three challenges. Here, we propose Learning to Design Optical‐Resonators (L2DO) to leverage RL that learns to autonomously inverse design nanophotonic laser cavities without any prior knowledge while retrieving unique design solutions. L2DO incorporates two different algorithms – Deep Q‐learning and Proximal Policy Optimization. We evaluate L2DO on two laser cavities: a long photonic crystal (PC) nanobeam and a PC nanobeam with an L3 cavity, both popular structures for semiconductor lasers. Trained for less than 152 hours on limited hardware resources, L2DO has improved state‐of‐the‐art results in the literature by over 2 orders of magnitude and obtained 10 times better performance than a human expert working the same task for over a month. L2DO first learned to meet the required maxima of Q ‐factors (>50 million) and then proceeded to optimize some additional good‐to‐have features (e.g., resonance frequency, modal volume). Compared with iterative human designs and inverse design via supervised learning, L2DO can achieve over two orders of magnitude higher sample‐efficiency without suffering from the three issues above. This work confirms the potential of deep RL algorithms to surpass human designs and marks a solid step towards a fully automated AI framework for photonics inverse design.