Traditional Approach to Materials Design and Process Optimization
The experimental design of new materials has traditionally been carried out through trial and error: a material is first synthesized based on previous knowledge and scientific intuition and is then tested to check whether it fulfills the requirements related to its application. In addition to material composition and topology, the synthesis also needs to be optimized, with many steps that can lead to failure and the need to iterate through materials or synthesis parameters until a solution is found. This approach requires important physical resources and time-demanding optimizations.
Machine Learning Optimizes Design of Experiment
Naturally, Design of Experiment (DoE) methodologies have been developed to systematically scan material synthesis parameters. However, such parameters are typically scanned one at a time to isolate their impact on the outcome, because this is how the human mind grasps the world. However, with the advent of machine learning (ML) methodologies over the last decade, this DoE approach can be optimized to further decrease the number of experiments needed to be performed before reaching optimal performance. Indeed, ML models benefit from making links between different sets of synthesis parameters in all dimensions of the possible space simultaneously.
Active Learning Methodology Developed at Matgenix (MGX)
The so-called Active Learning methodology consists in using ML to optimize the DoE and reach the best possible performances iteratively. In this approach, an initial set of experiments can be suggested where synthesis parameters are scanned semi-randomly (Latin Hypercube Sampling) as an ensemble rather than one dimension at a time. This allows for reducing the number of initial experiments to perform in the lab. Once those syntheses have been realized and data have been gathered, an ML model can be trained to model the synthesis outcome (outputs) based on the parameters (inputs). The ML model can then be used to provide suggestions of the next synthesis parameters to try in the lab: either those that maximize the performances (exploitation mode of active learning), those that will lead to the strongest improvement of the ML model (exploration mode of active learning), or a compromise between the two (through the acquisition function). Those suggestions are then realized in the lab, and performances are assessed. If a satisfactory solution has been reached, the process can stop. Otherwise, the new data can be used to feed the ML model, leading to better predictions and suggestions for the next round of this iterative active learning methodology.
Electrodeposition of Ni-W Films for Water Electrolysis: NICKEFFECT Use Case 1
In the NICKEFFECT project, Use Case 1 focuses on replacing Pt by Ni-W alloys for PEM water electrolysis. To that purpose, researchers at UAB quickly started to deposit Ni-W dense coatings through aqueous electrodeposition and measure their electrocatalytic activity (specifically, towards hydrogen evolution reaction, HER) in acidic media with electroanalytical techniques such as linear sweep voltammetry (LSV), searching for an efficient alternative to Pt in terms of cost and abundance.
Optimal Catalytic Performances Reached for Ni-W Films
After 24 depositions that exhibitted suboptimal HER performance, MGX trained an ML model on the data that had been collected by UAB to suggest the next deposition parameters to try. This required MGX to understand the data, automate the post-processing of the raw data from the equipment, and collaborate with UAB to define a score from the measured LSV curves that represents the overall performance of the coating, including the degradation of the samples. Suggestions of the next trials could be provided to UAB by using the ML model to find the set of deposition parameters that optimize this score. However, the first suggestions provided by the ML model did not lead to satisfactory performance. A second round of the active learning procedure took place with an improved model (including more data) and a refined score definition. New suggestions were provided, and after this second round of active learning, one of the new coatings (among 57 in total), deposited with parameters suggested by the ML model, showed very good electrocatalytic performance as well as low degradation with LSV cycles.
Success: Ni-W coatings can replace Pt for PEMWE
After repetition of the deposition and experimental characterization of the coating with the optimized parameters, it has been confirmed that Ni-W can replace Pt for PEMWE applications. This successful result, obtained through a strong collaboration between NICKEFFECT partners and published in open access in ChemSusChem (Wiley), has been the basis for further upscaling towards a viable industrialization.