Medica l Image Ana lysis & Ar t i f icia l Intel l igence Symposium 2022 Abstract: Accelerated Combinatorial Medical Material Development through Machine Learning from Theoretical and Experimental Data Sets Novel therapies and/or novel treatment methods require novel materials. Combinatorial Material Development is a strong approach if one is aiming for an accelerated material development for this applications. Both, theoretical and experimental attempts have been used in combinatorial material science [1]. The theoretical approach generally has higher degrees in freedom for performing the material “evaluation”. A cloud of calculations may use various small variations in the multidimensional parameter space, or let’s call it the condition vector (cA, cB, cC, T, t, pH,…) (with concentration of alloy constituting elements cA, cB, cC, temperature, time since preparation, pH of the solution,…) if that was the suggestion from the initial results. Experimental attempts are somewhat different. A material library with one or two lateral composition gradients is prepared in one batch and will be investigated typically in a way that some of the parameters are kept constant. All compositions will be usually studied at the same temperature; initial exposure time to air was identical for each alloy in the library; the pH of the solution will be the same during electrochemical experiments and will not change from one position to the next. These two approaches thus appear to be quite incompatible. However, through machine learning the complementary – and sometimes contradictory – data sets can be brought together to allow guiding each other. The challenge is to find a proper representation of the data that compensates for the misfit. Strict logic (PL1) will soon fail as it is not designed to bridge such mismatches. Defeasible reasoning such as in truth maintenance systems JTMS or ATMS are much better suited. Also, probabilistic approaches such as Bayes Networks can do a good job. Starting from the combinatorial experimental approach in a sophisticated research cluster the CALMAR system, multidimensional material libraries can be prepared and studied. Here, Al-Ho [2], Nb-Ti, CoCr and CoCrMo are taken as examples for which important material properties have been determined using EDX, XRD and electrochemical stability via f low type scanning droplet cell microscopy with downstream analytics (ICP-OES, ICP-MS) [3]. Regions of stability and relative predominance in thermodynamic Pourbaix diagrams have been computed using DFT for e.g. phosphate buffer solutions. Both data sets were then used for machine learning independently from each other to allow for a comparison of the predictions. Different artificial intelligence attempts are then used to try merging theoretical and experimental results.
RkJQdWJsaXNoZXIy NTA3MTM=