Research On High-dimensional Potential Energy Surface Fitting Based On Crossbar Array Architecture Of Non-volatile Memory Devices

Density functional theory(DFT),which won the Nobel Prize in chemistry in1998,is an important method for calculating atomic-scale microscopic processes.It is widely used in many scientific and technological fi elds,such as physics,chemistry,biology,pharmaceuticals,semiconductor design and so on.However,due to the high computational cost of DFT,the calculation of microscopic atomic motion based on DFT takes a very long time.For example,using a server wi th 8 CPU cores(specifically configured with:Intel CPU 2680V2),the parallel calculation of a 20picosecond(picosecond:10^-12 seconds)physical process of a benzene ring molecule with 12 atoms,and the DFT-based molecular dynamics simulation requires 37 h ours to run for 10,000 steps.Among them,more than 99%of the time is spent on high-dimensional potential energy surfaces(HD-PES)at the microscopic atomic scale.Therefore,in order to realize the efficient ca lculation of atomic-scale micro-processes in the fields of physics,chemistry,biology,pharmacy,semiconductor design and so on,it is necessary to study the efficient new computing technology of high-dimensional potential energy surface.To solve this problem,this paper uses DFT data to train the"integrated storage and calculation"crossbar array architecture based on new non-volatile devices to accurately fit and calculate the high-dimensional potential energy surface.By using a new neural network algorithm instead of the traditional DFT algorithm to calculate the high-dimensional potential energy surface,and by using the new architecture of"integration of storage and calculation"instead of the traditional architecture of"storage and calculation separation"of Central Processing Unit/Graphics Processing Unit(CPU/GPU),the calculation speed of atomic-scale microscopic processes is greatly improved on the premise of ensuring the calculation accuracy.The main innovations of this paper are as follows:1.In order to solve the problem that the computin g speed of the mainstream technology is too slow,a new high-speed computing architecture with high-dimensional potential energy surface is proposed.Based on the phase change heterostructure memory(PCH)and two-terminal single-poly floating-gate memristor(named Y-Flash),which are two new non-volatile devices rising in recent years,a new crossbar array architecture is designed,which increases the computing speed of high dimensional potential energy surface on microscopic a tomic scale by 6 orders of magnitude.Taking the calculation of the above benzene ring molecule as an example,the calculation time is reduced from 1.3 seconds of the current mainstream DFT single-step single-atom simulation(2.8GHz CPU)to 1 microsecond(with a 1 MHz frequency of designed architecture).2.In order to solve the problem of excessive calculation error caused by the performance fluctuation of mainstream non-volatile devices,a new high-precision calculation scheme of high-dimensional potential energy surface is proposed:multilayer perceptron(MLP)neural network and quantized neural network(QNN)based on vector-matrix multiplication,are combined with error control method to realize high-precision and reliable calculation of high-dimensional potential energy surfaces.The accuracy tests of nitrogen,water and benzene ring systems show that the proposed high-precision calculation method can achieve the fitting accuracy of about 50 me V/?(1 me V/?=1.6×10^-32 N),which is at the same level as the convergence residual of the mainstream DFT calculation method(about 40 me V/?),which effectively ensures the calculation accuracy.3.Based on the 180 nm CMOS process library and Cadence Virtusuo software,a series of designs such as operational amplifier circuit,current difference circui t and nonlinear activation function circuit are completed.Combined with the circuit model,the design and training of MLP and QNN are realized to fit the high-dimensional potential energy surface.Through the test and simulation,the high precision and hi gh speed described in the above two points are verified.The results of this paper are helpful to the design and development of new high-speed and high-precision computing techniques for atomic-scale microscopic processes.The related achievements are exp ected to be applied to physics,chemistry,biology,pharmacy,semiconductor design and other important scientific and technological fields.