Accelerated Prediction Of Packaging Materials Based On Machine Learning

As semiconductor technology develops,the packaging technology requires more and more microelectronic packaging materials.The traditional packaging material research strategy of experimenting with new materials or combinations of new elements under the guidance of theory and experience is no longer suitable for today's semiconductor and packaging technology,which is taking increased attention.At the same time,based on the high computational complexity of first principles and the increase of the explorable space of materials,the combination of first principles calculations and experiments is gradually unable to meet the demand for research on microelectronic packaging materials.Therefore,it becomes more and more important to propose a new research strategy that can efficiently explore in large material space for packaging materials.For the problems mentioned above,this work proposes a way to predict the configurational energy and elastic modulus of packaging materials based on machine learning and first principles to accelerate researchers' research on the performance of packaging materials.Firstly,based on Cu,the Cu Ni binary system alloy was established by substitutional doping of Cu with Ni in the way of multiple supercells.For this system,the cluster expansion theory is used to construct the features.By setting the cut of distance,the cluster correlation functions of some different orbits are taken as the machine learning features.Secondly,353 configuration energies and180 elastic modulus of Cu Ni are obtained by first-principles calculation,and a dataset is built.A feature filtering method is also proposed to filter machine learning features constructed based on cluster expansion theory using GA,RF and RFE to build feature sets based on different filtering schemes.Thirdly,four machine learning models,NN,GPR,RFR and KRR,are trained using feature sets and data sets to predict Cu Ni's configurational energy and elastic modulus.The results show that for different machine learning models with different screening methods,the number and combination of features required to achieve the optimal effect of the model are different.For configurational energy prediction,RF-NN scheme has the best performance,R2=0.95,MSE=4.44E-06ev/unit cell.The number of required features is 26.For the prediction of elastic modulus,RFE-RFR scheme has the best effect,R2=0.813,MSE=34GPa.The number of required features is 2.Furthermore,this thesis evaluates the correlation between configurational energy and elastic modulus from a machine learning perspective by using the configurational energy as a feature to predict the elastic modulus,and concludes that the correlation between the two is not significant.Finally,this thesis provides a preliminary exploration of the problem of little data for materials research based on transfer learning,using existing data and models,and proves to be effective and feasible.