The Self-Driving Lab: Turning Alloy Predictions into Experimental Knowledge

A promising composition still has to become a material

For a new jet-turbine alloy, predicting the composition is an early step. The elements may not mix as intended; different structures may emerge as the metal solidifies; casting, heat treatment and manufacturing from powder can change performance. Even a successful laboratory sample needs qualification before it can fly. Experiments establish what the proposed material can actually do.

That problem anchors Joseph Krause’s account of Radical AI in the Latent Space conversation. Introduced as the company’s CEO, Krause argues that the material itself supplies ground truth. Radical’s self-driving lab is intended to close a research loop: generate hypotheses, run experiments, capture results and return them to an AI scientist that chooses what to investigate next.

Krause attributes part of materials development’s slowness to fragmented institutions. Academia discovers candidates, smaller companies perform early tests, and larger manufacturers optimize established systems. Data and reasoning get lost between them. He recalls development estimates of 15–30 years, rather than a measured industry-wide timeline. Radical’s ambition is to preserve the connection through manufacturing so downstream outcomes can improve the next discovery campaign.

Following the material through the lab

Microstructure is the internal structure of a material, including the formations and phases that influence its properties. A proposed alloy may emerge from synthesis unevenly mixed or containing unexpected structures. Annealing and other thermal processing can change it again. Connecting a composition to performance therefore requires recording how the sample was made and what emerged at each stage.

Radical combines synthesis with characterization—determining what was actually made—and early property tests: - Oxidation testing examines behavior in oxidative or corrosive environments. - Tensile testing produces stress–strain curves. - Microindentation provides Vickers hardness and a company-developed indicator of whether a material is ductile. That ductility indicator is a proxy, not an exact measurement.

At the time of the conversation, the company captured discovery and testing data; manufacturing remained outside the loop. It assessed processability—whether a material could be formed or cast into the required shape—at roughly 200–500 grams. Ten-ton production was still beyond that work. Specialized wind-tunnel and torch tests went to third parties. The longer-term goal is a linked history from hypothesis through manufacturing and use.

Radical initially reports approximately 1,200 alloys made over five or six months, including 300 it describes as absent from prior literature. About ten showed especially promising performance, without accompanying quantitative benchmarks. Its focus includes high-entropy alloys, described here as five to seven elements in roughly equal atomic proportions, for applications involving extreme temperatures, pressure, irradiation or oxidation.

The application changes the objective

Faster discovery could enable concurrent engineering: developing a material alongside the product that needs it. A rocket booster, turbine or solar cell supplies requirements; experiments reveal attainable properties; material and product designs evolve together. The opportunity is to give established industries material options beyond their familiar alloys.

A successful early sample still faces qualification: producing material and running standardized tests to establish suitability for an application. Radical’s representative describes aerospace and military pathways as commonly taking about ten years. Recording where and why candidates fail could feed valuable constraints back into discovery.

The proposed change is to generate qualification evidence more effectively while preserving the safety standard. DARPA work exploring layer-by-layer qualification through additive manufacturing is offered as one approach, without a demonstrated time reduction. A bent phone and a failed turbine have different consequences. Automated manufacturing could supply more sensor data, but faster discovery alone does not establish flight safety.

Supply chains can change the objective from better performance to the same performance without a problematic element. Radical describes requests to remove hafnium from C-103, an aerospace alloy containing about 10% hafnium by weight, following a reported sharp price increase. The company says it achieved that substitution while meeting the same specifications; replacement composition, test results and qualification status are not disclosed.

Cost and component compatibility further constrain the design. A space application may tolerate expensive materials to obtain performance, while consumer electronics and medical devices face tighter economics. This explains Krause’s objection to treating alloy design as a representation of elements and bonds alone: industrial usefulness also depends on processing, microstructure, availability and the surrounding product.

Autonomy begins with a stuck metal puck

Human expertise enters the loop through annotations. A scientist examines a scanning electron microscopy (SEM) image—a picture of the material—and marks where dendritic formations appear. Radical calls this scientific intuition: expert interpretation that the AI scientist must learn to reproduce.

Instrument access is another prerequisite. Radical recalls vendors resisting software and data access roughly two years earlier, even for payment, before becoming more receptive. Its representative suggests a commercial explanation: proprietary analysis software helps distinguish instruments, so access to raw data can look like a threat to that value.

A host raises a practical objection: automated stations can still leave people carrying trays between machines, and partial automation can slow the overall workflow. Radical distinguishes an automated lab that executes experiments from a self-driving lab that directs research campaigns. The latter must choose and carry out the route toward a research objective.

That route includes stubborn physical problems. Synthesized alloys emerge as small pucks called buttons, which can stick to their trays after heating. Radical built custom actuators to free and manipulate them without damaging the samples or disturbing their microstructure. An X-ray diffraction instrument, or XRD, presents a different challenge: its awkward geometry and rigid sample mount require a custom gripper. Scientists use the resulting diffraction spectra to help identify material phases.

A lab operating system coordinates the work. It tracks sample identity, determines which instrument should receive each sample, and combines vision with instrument and laboratory sensors. Post-synthesis quality checks can stop an unsuitable experiment before it consumes characterization time. Robots unload samples and transfer them between stations, a capability Radical reports already operating.

The infrastructure is specialized for alloys. Some characterization tools can serve other materials, but the oxidation chamber targets a particular customer application, and the synthesis tool does not handle ceramics or polymers. Radical originally envisioned seven labs for seven material systems. Customer questions about missing tests and scaling to 300 pounds shifted the priority toward vertical integration before expansion.

Campaigns are the unit of learning

Semiconductor interconnects offer another application target. Radical reports investigating replacements for materials used in back-end-of-line integration, where it sees an energy-efficiency bottleneck. Its representative estimates potential gains of 2–5×, and speculates about larger gains later, without defining an efficiency metric or baseline. Barrier layers, interfaces and new testing capabilities complicate integration into commercial chips.

For some defense and space alloy applications, the company considers three to five years plausible. That outlook explicitly excludes comparable expectations for crewed-flight jet turbines, where human-safety constraints make deployment harder.

Within the lab, Radical reports an active-learning loop: experimental results inform the next experiments. It works campaign by campaign: 1. The AI scientist selects a set of materials to make and test. 2. The lab carries out synthesis, characterization and early property testing, extracting characterization data automatically. 3. Machine-learning models and human scientists analyze the results. 4. Results enter a database used to design subsequent campaigns. Manufacturing outcomes were not yet included.

Hypotheses change between campaigns, after several outcomes provide enough evidence to revise them. The representative estimates capacity for seven to ten concurrent campaigns across different systems, with updates daily or every other day as results arrive.

Automation remained uneven. Metallurgy PhDs still operated synthesis equipment, visually identifying unmelted regions and directing a plasma torch toward them. Models were learning that control task, without a reported human-comparison benchmark. Characterization loading, unloading and operation were automated, while scientists continued annotating results. Microindentation and oxidation tests were automated; tensile testing was nearly automated.

More experiments, or better scientific judgment?

Radical says its AI scientist explored alloy families absent from the published literature it could access. Some candidates could be synthesized despite scientists’ expectations that elements would evaporate, fail to cast or produce unsuitable microstructures. These are examples of challenging assumptions; no compositions or measured properties accompany them.

A host asks the essential counterfactual: would human scientists with the same throughput and freedom explore as effectively? Radical acknowledges that literature retrieval sometimes draws the AI toward familiar approaches. Higher throughput makes uncertain experiments more affordable. Its representative recalls roughly 50 experiments a year during a PhD, with fabrication, synthesis and testing taking about two weeks at a time.

Radical’s reported rate was eight to twenty experiments per day across the entire lab. Harder-to-cast refractory alloys pushed throughput toward eight; easier systems involving titanium or aluminum allowed rates nearer twenty. Reported costs ranged from roughly $60 to $300 per experiment depending on the elements, with accounting boundaries unspecified. Reaching 100 experiments per day around June or July was a rough future target.

The broader AI argument is that literature, images, hypotheses and property measurements can be connected in parallel. The host presses again: this is still a small-data setting, where a scientist might solve a local problem before a model becomes useful. Radical’s answer combines AI screening of a large combinatorial space with scarce experimental evidence. Each alloy reportedly yields roughly 50–150 data points, depending on the images and spectra collected; the company acknowledges that the dataset remains small.

The comparison with DARPA/GE MACH does not establish a clean speedup. Radical’s representative describes MACH as producing 500 alloys in about a year, then gives three months for Radical’s 1,200 alloys, differing from the earlier five-to-six-month account. The companion article uses six months and calls MACH’s figure a target. Completing 500 alloys in five business days remains Radical’s goal.

A materials data bank needs more than structures

Radical identifies experimental feedback as its main bottleneck: running enough physical experiments to change the AI scientist’s choices matters more than adding compute. It wants a materials counterpart to the Protein Data Bank, encompassing the properties needed to judge industrial usefulness as well as crystallographic structures.

Narrow modeling tasks can still deliver major advances. Radical uses segmentation models to identify features in SEM images, such as dendrites, cracks and defects. Relating those features to strength is a useful task in its own right. Predicting microstructure or crack propagation, however, does not establish whether the material can be made into powder, cast successfully or manufactured into a product.

Scale-up can reveal failure modes that researchers did not know to measure, so the dataset must evolve with the process. Radical’s representative advocates a correspondingly demanding definition of discovery: design, synthesis and characterization are milestones; a scaled material succeeds when it reaches a product.

An advisor with 35 years at 3M illustrates the missing manufacturing knowledge: an experienced operator may know exactly when to turn a knob and how far, without an explicit, repeatable formula. Radical has not solved how to capture that judgment. Its proposed approach combines expert input with sensors and automation. Nearer term, it seeks manufacturing partners who can explain expected material behavior alongside supply-chain, cost and application constraints.

Building the infrastructure and the workforce

An early instrument required a two-week software sprint just to obtain programmatic control. Building the lab also demanded more specialties than expected: computational and experimental materials science, mechanical engineering, mechatronics, full-stack software and applied machine learning. Off-the-shelf robotics left substantial work in path planning, perception and custom manipulation.

Several improvements make that integration easier. Machine-learned interatomic potentials help accelerate simulations and narrow candidates before experiments. Cheaper robots, better actuators and custom grippers expand manipulation options. Instrument vendors increasingly support software interfaces. Radical’s representative calls hardware and infrastructure the biggest change, while acknowledging that the field still needs a major result to persuade cautious customers.

Radical advocates self-driving labs as national infrastructure. Its proposed public-private partnership combines national-lab computing, scientists, instruments and experimental data with private-sector integration capabilities. The competitive argument is to improve research systems and workforce output. A host notes that China can adopt the same productivity tools, which the representative accepts.

Higher throughput does not eliminate physical feedback times. Some experiments take weeks or years. A hypothetical facility with a thousand microscopy or diffraction instruments illustrates one route to larger datasets, at substantial infrastructure cost. The more immediate request is for vendors to build tools for agents and robots, allowing scientists to operate integrated research systems rather than each instrument separately.

The career advice is to bring your specialization into science. An ML engineer may recognize a segmentation task where a metallurgist first sees dendritic formations and grain boundaries. Materials scientists should learn to use AI tools effectively without assuming they must retrain as ML engineers. The collaboration needs both kinds of expertise.

Inside the AI scientist

Radical’s AI scientist is a multi-agent system. An orchestrator proposes hypotheses, which undergo an undisclosed internal assessment before reaching the lab. Supporting inputs include a custom literature-review agent that extracts relevant figures and information, past experimental results, and paid CALPHAD data. CALPHAD’s methodology is not explained in the conversation; its specified role is an additional scientific data source for campaign design.

For experimental interpretation, Radical describes a Qwen-based vision-language model fine-tuned to extract scientific knowledge from laboratory images and data. The conversation calls it MATRIX; the companion article distinguishes MATRIX, the open dataset and benchmark, from MATRIX-PT, the model based on it. The MATRIX/MATRIX-PT resources are available through Radical’s Hugging Face organization.

Radical reports that training on experimental interpretation improved broader scientific reasoning as well as data reading. The representative recalls roughly 5–16% improvement, without specifying evaluation tasks, baseline or whether that means relative gains or percentage points. Transfer to biology and mathematics as an exception are also tentative recollections. The proposed mechanism is that learning to interpret experimental evidence builds capabilities useful beyond the immediate reading task.

A customer demo makes accumulated knowledge inspectable: a generated hypothesis links to the earlier experiments that informed it. Specialized agents and models are intended to combine retrieval, interpretation and hypothesis generation, giving each new campaign an experimental basis that users can examine.

Open models, experimental advantage

Radical gives three reasons for open sourcing: encourage adoption, learn from a community larger than its internal team, and concentrate its competitive advantage in experimentation. [TorchSim](https://github.com/torchsim/torch-sim) is one example. The companion article describes it as a PyTorch-based molecular-dynamics simulation framework; Radical reports spinning it into a nonprofit to continue development with the community.

Releasing a model does not release all the lab’s data. Radical says its public MATRIX work used public data, while its internal work incorporates proprietary experimental data. The company welcomes better external models and combines custom scientific models with outside systems, including paid large language models such as ChatGPT and Claude. It says it neither sells models nor builds its own general-purpose language models.

Experiments are the moat is Radical’s business thesis: shared model improvements can strengthen a system whose distinctive capability is producing and learning from physical evidence. Full laboratory autonomy remains a future milestone. The conversation closes with the guest proposing a return episode when the lab reaches it.