SparksMatter inorganic design, AHOIS encoding hypothesis, RL quantum calibration

An evaluation of the SparksMatter AI model, which demonstrated high performance in designing novel inorganic materials. Other developments include AHOIS, a multi-agent system that autonomously proposed and validated a physical hypothesis on an optical platform. In quantum computing, a new framework unifies reinforcement learning with quantum error correction to enable continuous self-calibration, achieving record logical error rates on a superconducting processor.

AI Agents for Materials Discovery and Scientific Autonomy

SparksMatter, a multi-agent AI model, exhibited high performance in inorganic materials design. It generated novel, stable inorganic structures tailored to user specifications across thermoelectric, semiconductor, and perovskite oxide materials. Benchmarking indicated SparksMatter scored higher in relevance, novelty, and scientific rigor compared to other models. This system generates ideas, designs experimental workflows, executes them, and refines results, also critiquing its own responses and suggesting validation steps.
AHOIS, a multi-agent AI scientist, autonomously discovered and validated a random-interference encoding hypothesis during evaluation on a multimode-fibre optical platform. AHOIS identified task-adaptive sparse-measurement strategies and diagnosed failure modes. It also translated a published imaging protocol into an executable workflow. This system integrates Socratic methods into closed-loop experimentation to construct, challenge, and revise physical explanations based on evidence. Ablation studies showed Socratic interrogation improved physical consistency, hypothesis completeness, uncertainty calibration, and experimental-plan validity within AHOIS. AHOIS's discovered encoding achieved 76.97% accuracy on MNIST.

Reinforcement Learning in Quantum Computing

Researchers integrated reinforcement learning (RL) with quantum error correction (QEC) to enable continuous self-calibration during quantum computation. QEC's error-detection events serve as a learning signal for an RL agent, which continuously adjusts control parameters to stabilize the quantum system without interrupting computation.
The RL-QEC framework was experimentally demonstrated on a Willow superconducting processor. This implementation led to record logical error rates for surface codes at 7.72(9) × 10−4 per cycle and color codes at 8.19(14) × 10−3 per cycle. The logical stability of the surface code improved by 3.5 times against injected drift. Numerical simulations confirmed the framework's scalability for large QEC codes, with optimization speed remaining constant for systems with tens of thousands of control parameters.
A reinforcement learning-based quantum compiler optimized two-qubit circuits on a superconducting processor, surpassing conventional methods by discovering hardware-amenable circuits with near-optimal lengths. Researchers further developed a variational reinforcement learning-based quantum compiler that consistently identified near-optimal circuits under hardware constraints. However, a reinforcement learning-based quantum compiler did not achieve high fidelity for three-qubit circuits. Researchers also analyzed the impact of decoherence and gate errors on reinforcement learning-based quantum compilers.

Sources

Autonomous in-silico inorganic materials discovery via multi-agent physics-aware scientific reasoning
Conventional machine learning approaches accelerate in-silico inorganic materials design via accurate property prediction and targeted material generation, yet they operate as single-shot models limited by the latent knowledge baked into their training data. A central challenge lies in creating an intelligent system capable of autonomously executing the full in-silico inorganic materials discovery cycle, from ideation and planning to experimentation and iterative refinement. We introduce SparksMatter, a multi-agent AI model for automated inorganic materials design that addresses user queries…
Socratic agents for autonomous scientific discovery in high-dimensional physical systems
The automation of scientific discovery has reached an inflection point. While AI systems now operate instruments, optimize parameters and generate hypotheses, most remain procedural: they execute workflows fixed by human designers. True autonomous science demands epistemic autonomy--the capacity to construct, challenge and revise physical explanations in response to evidence. Here we introduce AHOIS, a multi-agent AI scientist that embeds Socratic midwifery into closed-loop experimentation. A physics-critic agent interrogates hypotheses through causal questioning, constraint checking,…
Reinforcement learning control of quantum error correction
Quantum error correction (QEC) is the primary strategy for protecting a quantum computer from the environment1,2. The prerequisite of QEC is that errors must remain sufficiently rare, which requires perpetually adapting the control parameters of the computer to the drifting environmental conditions. The current solution to this problem is to terminate the entire quantum computation for recalibration, but it is incompatible with the long runtimes of future quantum algorithms3,4. Here we address this challenge by unifying calibration with computation. We grant the QEC process5–11 a dual role:…
Quantum compiling with reinforcement learning on a superconducting processor
Abstract Effectively implementing quantum algorithms on noisy intermediate-scale quantum (NISQ) processors is a central task in modern quantum technology. NISQ processors feature tens to a few hundreds of noisy qubits with limited coherence times and gate operations with errors, so NISQ algorithms naturally require employing circuits of short lengths via quantum compilation. Here, we evaluate a reinforcement learning (RL)-based quantum compiler on a superconducting processor. Our experiments reveal that for two-qubit circuits, the RL-based compiler surpasses conventional methods,…

Full transcript

What if an AI could not just predict material properties, but autonomously design entirely new ones? This is the changing landscape of scientific discovery we track on Agents in Research. We'll start with new systems for materials science. We're seeing new AI systems that are moving beyond data analysis and into the realm of autonomous scientific discovery. Right, from being a tool to more of a collaborator. And this is happening in very different fields, from materials science to quantum computing. Let's start with materials science. A multi-agent AI model called SparksMatter has been introduced for designing inorganic materials automatically. So it's not just sorting through known materials, it's creating new ones? How does that work? It generates ideas, designs experimental workflows to test them, and then refines the results based on what a user is looking for. It was evaluated on thermoelectric materials and perovskite oxides, for example. And what was the outcome? Did it produce anything viable? It generated novel and stable inorganic structures that met the specified objectives. In benchmark tests, it received high scores for scientific rigor and relevance, and especially high scores for novelty. So it's not just recombining things in predictable ways. It's finding genuinely new configurations. There's another system that seems to push this idea of rigor even further. That would be AHOIS, which is for autonomous discovery in physical systems. Its distinct feature is what it calls a 'physics-critic' agent. A critic agent? So it's designed to argue with itself? In a structured way, yes. It uses a Socratic method. The critic agent interrogates the system’s own hypotheses with causal questions and actively tries to falsify them. That sounds like a way to build scientific skepticism directly into the process, to avoid generating plausible but incorrect theories. And the approach seems effective. In one evaluation, AHOIS autonomously proposed and then validated a hypothesis about random-interference encoding. Ablation studies showed this Socratic part was key to the consistency of its hypotheses. And that encoding it discovered wasn't just a theoretical exercise. It was tested on the MNIST dataset and achieved nearly 77% accuracy. This idea of AI agents managing complex processes also appears in quantum computing, starting with compilers. Which are essential for translating algorithms into physical operations on the qubits. A notoriously hard problem. One approach used a reinforcement learning-based compiler. It surpassed conventional methods, but it had a clear limitation: it worked for two-qubit circuits, but its fidelity was low for three-qubit circuits. That's a familiar scaling challenge. The complexity grows quickly with each additional qubit. But a newer, variational reinforcement learning-based compiler was developed to handle that. It can consistently identify near-optimal circuits, even when dealing with hardware constraints. So the AI learns to navigate the specific flaws and limitations of the physical hardware. That seems related to the larger problem of quantum error correction. It is. And on that front, researchers developed a system that unifies error correction with continuous calibration. It uses the error-detection events from the correction process as a learning signal for a reinforcement learning agent. That's an elegant solution. It turns the problem—the errors—into the very signal that teaches the system how to stabilize itself, and it does this without interrupting the main computation. When demonstrated on a superconducting processor, it achieved record-low logical error rates. For the surface code, the average logical error per cycle was 7.72 times 10 to the negative 4. It also improved the code's stability against injected drift by a factor of 3.5. Those are specific performance gains, but the scalability might be the bigger story. Numerical simulations suggest the framework's optimization speed stays constant, even as the system grows much larger. Correct. The findings suggest it could be applied to systems with tens of thousands of control parameters, indicating it's a scalable path forward. That is all for today's report. We'll have more updates for you next time. From Agents in Research, thanks for listening.