Here we introduce a new paradigm in PHIN-OS, composable design space search tasks that enables using state of the art Bayesian Optimization (BO) in any atomic scale simulation workflow. The functionality reduces the cost of materials design by 97% by efficiently navigating vast design spaces. Two case studies—optimizing molecular liquid density and identifying low-energy crystalline interfaces—demonstrate BO's superiority in finding optimal results in a small fraction of the total search space compared to brute force or random sampling. The native implementation democratizes advanced search techniques, allowing materials scientists to easily transition to guided, data-driven exploration without needing specialized programming expertise or managing bespoke execution pipelines.
Finding novel materials with tailored properties is bottlenecked by the vastness of the design space. In atomic-scale simulation experiments, optimizing a material's performance often requires exploring a high-dimensional parameter landscape—be it composition ratios or structural configurations. Manually searching this space is not only time-consuming but often impractical, as the number of possible combinations can easily exceed millions.
This challenge makes automated design space search critical. By leveraging intelligent algorithms, we can move beyond exhaustive or pseudo-random sampling and focus computational resources on the most promising candidates. Central to automated optimization is Bayesian Optimization (BO). Unlike simple random sampling, BO uses the results from previous simulations to build a statistical surrogate model of the relationship between input parameters and the desired material property. This model guides the selection of the next greatest expected utility experiment, systematically and efficiently zeroing in on the optimum. For materials design, BO is essential for accelerating discovery and minimizing the required—and often expensive—simulation time.
Integrating Bayesian optimization (BO) into production environments presents significant challenges, despite its conceptual simplicity. The primary hurdle is the often manual and bespoke nature of creating active learning pipelines. This typically involves developing custom scripts for each experiment to interface with specific simulation code, along with managing parameter input/output parsing and model updates. The reliance on custom development sets a high barrier to entry, requiring specialized programming skills and consuming substantial development time for every new simulation setup. To overcome this limitation and make advanced search techniques widely accessible, we have natively integrated Bayesian Optimization into PHIN-OS. This implementation democratizes the use of BO by abstracting away the complexities of data, model, and orchestration management, allowing users to concentrate on the learning process. Crucially, the BO layer can be composed on top of any existing simulation workflow without requiring modification, fundamentally transforming how materials scientists approach design space exploration.
To showcase the power and efficiency of Bayesian Optimization in this context, we present two distinct simulation challenges and compare the performance of a standard random search controller against a BO controller.
The first experiment pertains to a common experiment setup where the simulation workflow has hyperparameters that can be tuned to improve the desired result. Here we focus on optimizing the density of a molecular liquid. The mixture contains three different molecules, Diethyl Carbonate (DEC), Ethylene Carbonate (EC), and Acetonitrile (ACN). These three molecules are the backbone of modern liquid electrolytes for high performance lithium-ion batteries, and determining their optimal concentration to optimize various mixture properties such as density, viscosity, freezing temperature, and more can improve the performance and reduce the cost of batteries substantially.
Here we narrow down an experiment workflow that computes the density of a liquid using PHIN’s no-code task provisioning. We wrap this within a Design Space Search object to be able to access all the experiment hyperparameters. The setup of this is seen in Figure 1. The individual tasks of the experiment create the atomic structures of the input molecules that a mixture generator uses to create a molecular liquid. The structure and volume are relaxed to find the minimum energy configuration to compute the density. This density, a scalar, is the objective value that the design space search uses to predict the next molecular liquid to test.
After the simulation workflow is set up, any of the parameters can be accessed for use in the design space search. As a demonstration, we choose the relative fractions of the molecules and enumerate design spaces for each molecular fraction. Here the design space is an enumerated type where we specify the specific values chosen for the allowable relative fractions. Combinatorially searching the design space enumerated by the three design variables gives a total of 441 parameter sets in the design space.
Traditionally, finding the optimal composition within this complex design space has required a manual, iterative development process. This approach involves searching for the non-linear relationship among the three fractions, often resulting in a pseudo-random exploration of over 400 possible combinations. By contrast, with PHIN-OS managing the optimization, we simply select a few starting points in the design space and allow the system to take over.
In Figure 2, we see the power of Bayesian optimization. In just 15 iterations, (roughly 3% of the total design space), the Bayesian controller is able to lower the mixture density much faster than the random search and even find a completely lower energy configuration within the design space. This shows the Bayesian controller’s ability to learn the high dimensional, non-linear relationships between the concentrations of the different mixtures, allowing for accelerating the search for the best configurations by 30x+, representing very large performance and cost savings.
The second experiment pertains to another common experiment setup where instead of changing a parameter of the experiment, the input structure itself is changed. This is common in atomic scale simulations where the lowest energy structure of a list is needed. Take for example finding the lowest energy interfacial structure between two crystals, for a common interfacial structure of silicon - silicon-dioxide, there can be hundreds of interfaces. Finding the lowest energy of a list of structures is a common occurrence in atomic scale simulations, where one might as likely need to find the lowest energy crystal structure or vacancy type. With PHIN-OS, it is as simple as creating a list of different input structures and having PHIN-OS automate the process of finding the optimum.
Here we focus on optimizing the formation energy of a silicon - silicon dioxide interface. And showcase the power of PHIN-OS by generating 100 different types of interfaces with maximum miller-indices of 1. All the candidate interfaces are shown in Figure 3.
The structures are enumerated as different inputs and can be easily selected as a controlled parameter just like any other simulation parameter. In Figure 4, the experiment for calculating the formation energy of a single structure is shown, which is composed of just 3 tasks. A minimization simulation task that finds the minimum energy structure of the input interface and two processing tasks. Once set up, PHIN-OS orchestrates the design space search to find the lowest energy structure.
The different types of structures in Figure 3, represent a high dimensional, highly non-linear design space that has to be optimized over. In these scenarios, engineers typically fall back on brute force techniques that sample the entire design space to find the optimum value. In Figure 5, however, we show how Bayesian optimization is able to find the best structure in 10 iterations, a 90% reduction in cost compared to the entire design space. This dramatic improvement of the Bayesian approach to Random or brute force approaches highlights the value of having Bayesian methods natively accessible in every simulation workflow.
The integration of Bayesian Optimization (BO) directly within the PHIN-OS platform represents a paradigm shift in computational materials design. The experimental results clearly demonstrate that native BO orchestration dramatically accelerates the search for optimal material properties, whether tuning mixture compositions or identifying low-energy structural configurations. By consistently outperforming standard random search by orders of magnitude this work showcases the value of utilizing advanced machine learning techniques to reduce computational time and cost.
The true impact, however, lies in the democratization of advanced search methodologies. By abstracting the complex coding and data pipeline management typically associated with active learning, PHIN-OS lowers the barrier to entry for materials scientists. Researchers can now seamlessly deploy state-of-the-art optimization techniques without requiring specialized expertise in machine learning or bespoke software development. This fundamentally changes the workflow from a manual, iterative trial-and-error process to a guided, data-driven exploration, making digitally driven materials development accessible and practical for any materials scientist.
