With the pressing need to decarbonize energy production, research is being accelerated in the field of nuclear power generation as a potential source for clean energy. Among the various structural materials used in a nuclear power reactor, Alloy 690 is used widely to construct and repair pressurized water reactor components. Since the adoption of alloy 690 in the nuclear industry, extensive research has been conducted on developing suitable welding procedures and filler metals that can produce weldments with the superior structural integrity required for service conditions. Extensive research on Ni-Cr-Fe base filler metals have been carried out for welding alloy 690 and has resulted in the development of filler metal class 52 (AWS A5.14). Since the inception of filler metal 52, improvements to filler metal composition have been conducted by the addition and control of B, Zr (52M), Mo, and Nb (52MSS) to improve the solidification cracking (SC) and ductility dip cracking (DDC) resistance. Recent research has introduced a new class of filler metal with superior SC and DDC resistance, 52MSS-Ta which uses Ta additions to replace Nb. While the incremental improvements in filler metal design over the decades have been through gradual changes in filler metal composition by varying one element at a time, an accelerated filler metal design approach that can find the optimum composition is needed.
Here we present an approach involving a combined high throughput CALPHAD data generation, data visualization, machine learning, synthetic data generation, and data mining techniques to accelerate the design of Ni-Cr filler metals for nuclear power generation. Firstly, several random compositions within the filler metal range were generated and Scheil solidification simulations were performed for all the randomly generated compositions. Following Scheil calculations, MX and Laves phase fractions, solidification range, crack susceptibility coefficient, and Kou index for solidification cracking were extracted for all the compositions. Using our previous experimental data on strain to fracture (STF), synthetic data was generated to train a machine learning model to predict STF. Finally, 1,000,000 synthetic compositions along with the corresponding MX and, Laves fractions, solidification range, crack susceptibility coefficient, and Kou index were generated for re-evaluating solidification cracking and DDC susceptibilities using machine learning and data mining.
Results and Discussions:
From the coupled approach to alloy design, it was found that Mo, Ta, Cu, Nb, Cr, and C elements increase the strain to fracture values whereas Mn, Al, and Ti decrease the strain to fracture resistance, indicating the roles of individual alloy elements in the filler metal on DDC. Upon generating synthetic data for STF and developing regression models for predicting STF, it was found that predictions from the gradient boosting regression algorithm were more accurate with an R2 of 0.94. Upon generating 1,000,000 synthetic compositions, it was found that C, and Nb mainly govern the fraction of MX. Nb, and Fe contents govern the fraction of Laves phase. Solidification cracking was governed mainly by the amounts of Si, Mo, Al, and Nb, and strain to fracture values were principally governed by the amounts of Mo, Nb, and Ta. Data mining of the 1,000,000 synthetically generated compositions resulted in optimized compositions with STF > 25 % higher than the existing filler metal compositions. Experimental testing is ongoing to validate these modeling results.
The established machine learning and data mining approach shows high potential for further improving SC and DDC resistance of 52MSS-Ta. Multiple key indices of SC and DDC were considered and analyzed with large sets of CALPHAD data and synthetic data. Data mining was used to optimize chemical compositions for meeting the design criteria of crack-free Ni-Cr filler metal. This approach can also be applied to accelerate other alloys and materials design with high efficiency and high accuracy.