Pure metals like aluminum or titanium don't always have the desired material properties - strength, hardness, ductility or corrosion resistance - for a given application. For this reason, researchers seek out novel alloy solutions, mixing a primary metal element with a series of other elements to create a material with tailored properties for uses in aerospace, defense, automobiles, energy and more.
But discovering a new alloy isn't just a matter of mixing a few elements together: it is a precise science that can take years of trial and error.
In an ambitious venture to cut down alloy development times from years to months, materials scientists and engineers at Lawrence Livermore National Laboratory (LLNL) are partnering with Cornell University to build the Autonomous Alloy Prediction and EXperimentation (APEX) platform for 3D printing, grinding, polishing and characterizing alloy samples.
Funded by LLNL's Laboratory Directed Research and Development program, this platform will combine industry-matured robotics and automation technologies with cutting-edge machine learning capabilities, turning a traditionally tedious and repetitive research process into a fully automated cycle.
"Our end goal is to make APEX the first self-driving laboratory for alloy discovery at LLNL, capable of working around the clock to collect experimental data and autonomously design, build and test novel alloys," said Mason Sage, APEX principal investigator and LLNL robotics engineer.
Working through challenges
Researching new alloys is complex for several reasons, but primarily because the design space expands exponentially with each added processing or compositional variable.
"Imagine you're baking cookies and want to create the perfect recipe, so you decide to experiment with three ingredients: sugar, butter and flour," Sage said. "If you test each ingredient with 10 different amounts, you'd need to bake 1,000 cookies to try every possible combination (10 sugar variations × 10 butter variations × 10 flour variations). Now, imagine you have 20 or 30 ingredients to test, suddenly, the number of cookies you'd need to bake becomes astronomically high - somewhere around 100 quintillion combinations, or 100 with eighteen zeros following it.
"This is the challenge before us. In theory, there hasn't been enough time since the universe began to test every possible combination - even if we conducted an experiment every second."
To navigate this enormous and complex parameter space, APEX leverages robotics and machine learning to accelerate sample fabrication, preparation and inspection by intelligently exploring different combinations.
"Our goal is to use machine learning to analyze the information we have collected from all our processes, learn from past experiments and design new experiments," Sage said.
After selecting an alloy to explore, the APEX platform builds the metal sample layer-by-layer using an additive manufacturing method called directed-energy deposition, a type of 3D printing that uses a high-power laser to melt and fuse metal powders onto a substrate.
Then, to prepare a printed sample to be microstructurally and mechanically characterized, its surface must undergo grinding and polishing. Traditionally, these steps can take two to three days to complete by hand, but APEX is expected to simultaneously process multiple samples at a time, producing dozens of samples a day.
Once a sample is prepped, APEX can evaluate its properties through a series of characterization tests, including examining its surface under a microscope, indenting it with a diamond to measure hardness and crushing it to assess its response to compression.
Finding ways to automate these labor-intensive steps has been a central challenge for the APEX team, one that is critical to the project's success.
"Trying to not only determine what is really important to our research process but also how to turn that into something you can control a robot to do has been an interesting undertaking," said LLNL research scientist Michael Juhasz.
For example, when a researcher grinds a sample, they do so by feel, tuning into the vibrational patterns of the machine and sample to ensure they are applying the right amount of pressure. This left the team wondering: How do you relay this highly developed skill to a robotic system? One creative solution the team came up with was to strap a vibration sensor to the machine to capture these "feelings" numerically.
"The sensor produces data in real time, allowing us to put numbers and parameters to the different vibrations, which can guide APEX when it's grinding a sample," said LLNL research scientist Alex Baker. "This project is really trying to understand where the automation complexity meets the material science complexity and then translating between the two."
Closing the loop
From sample fabrication to grinding and polishing to characterization, APEX works to collect data at every phase, generating a full history for each sample that passes through the platform. In the future, APEX aims to integrate this data with the Materials Acceleration Platform (MAP), combining physics-based models with machine learning to help design and improve the next round of APEX-produced samples.
The MAP is a framework for integrating computer models that uses algorithms to design optimal materials under different constraints (limitations related to how material properties interact and impact one another) in an effort to find an ideal composition that embodies all desired properties. This is particularly important because optimizing materials often involves trade-offs, where improving one property can sometimes lead to compromises in another. MAP helps strike a balance to achieve the best possible combination of properties.
While one model might excel at predicting a material's hardness, another might be better suited for predicting its ductility.
"MAP connects all of these separate, siloed models and weaves them together to help us predict new materials that we haven't tested yet," said LLNL computational scientist Brandon Bocklund. "The vision of the project is to really close the loop of materials development and acceleration - letting the system learn and update the models autonomously."
Currently in the early stages of its development, APEX is being trained to work with stainless steel alloys as the research team proves its feasibility.
"We chose stainless steel because it is very a well characterized alloy," Juhasz said. "This makes it the ideal control, allowing us to ground what APEX is doing and perfect the system before diving deeper into our materials development research."
The plan is to make APEX as extensible as possible by its anticipated completion in 2027, allowing researchers to adapt the platform to accommodate a variety of unique scientific challenges and incorporate new characterization or diagnostic techniques as needed.
This one-of-a-kind project exemplifies LLNL's multidisciplinary approach to pushing the boundaries of scientific discovery, demonstrating that when experts from different fields come together, no problem is too "unsolvable."
"It's been great seeing the synergies between deep scientific knowledge of the materials experts meshing with the technical expertise of the engineers," Sage said.
Baker echoed this sentiment, "There have been many instances where we each could see the subject matter expertise the other side of the project was bringing."
-Shelby Conn
