13th symposium on Discovery, Fusion, Creation of New Knowledge by Multidisciplinary Computational Sciences


Can deep learning replace current numerical weather prediction models?
DURRAN Dale,  (Atmospheric Sciences, Univ of Washington) WEYN Jonathan, CARUANA Rich (Microsoft), CRESSWELL-CLAY Nathaniel (Atmospheric Sciences, Univ of Washington)

We present a  data-driven global weather-forecasting framework using a deep convolutional neural network (CNN) to forecast six key variables that depend on horizontal position and time.   The variables are carried on a cubed sphere, which is a natural architecture on which to evaluate CNN stencils. In addition to the forecast fields, three external fields are specified: a land-sea mask, topographic height, and top-of-atmosphere insolation which varies as a function of spatial location, time of day, and calendar date. The model is recursively stepped forward in 12-hour time steps while representing the atmospheric fields with 6-hour temporal and roughly 1.4 x 1.4 degree spatial resolution.
The extreme computational efficiency of our Deep Learning Weather Prediction (DLWP) model allows us to create an ensemble prediction system requiring just three minutes on a single GPU to produce a 320-member set of six-week forecasts.  Ensemble spread is primarily produced by randomizing the training process to create a set of 32 DLWP models with slightly different CNN filter coefficients. 
Although our DLWP model does not forecast precipitation, it does forecast total column water vapor, and gives a reasonable 4.5-day deterministic forecast of Hurricane Irma. In addition to simulating mid-latitude weather systems, it spontaneously generates tropical cyclones in a one-year free-running simulation.  Averaged globally and over a two-year test set, the ensemble mean RMSE retains skill relative to climatology beyond two-weeks, with anomaly correlation coefficients remaining above 0.6 through six days.
Our primary application is to subseasonal-to-seasonal (S2S) forecasting at lead times from two to six weeks. Current forecast systems have low skill in predicting one- or 2-week-average weather patterns at S2S time scales. Probabilistic skill scores show our system performs  similar to the European Centre for Medium Range Weather Forecasts (ECMWF) S2S ensemble at lead times of 4 and 5-6 weeks. At shorter lead times, the much more computationally demanding ECMWF ensemble performs better than DLWP.
Development and application of real-time time-dependent density functional theory (RT-TDDFT) code, INQ, optimized for hybrid CPU-GPU HPC systems
Tadashi Ogitsu (Lawrence Livermore National Laboratory) 

The emergence of Density Functional Theory  (DFT)  together  with breakthroughs in algorithms and rapid increases in computer performance contributed significantly to the success of modern electronic structure theory.  While DFT has been being widely used by scientists and engineers as parameter free atomistic electronic structure simulation tool due to its practicality, eg. good balance between accuracy and affordable computational cost, time dependent counter part to DFT, time dependent DFT (TDDFT) method is yet to be mature enough to be a practical tool.
In this presentation, we will introduce the open-source real-time time-dependent density-functional-theory (RT-TDDFT) code, named INQ, being developed under DOE Computational Materials Science Software Center for Nonperturbative Studies of Functional Materials Under Nonequilibrium Conditions. INQ code takes a modular design and being optimized for CPU+GPU hybrid HPC systems such as Sierra at Lawrence Livermore National Laboratory. The software design prioritizes the code portability for future HPC systems such as El Capitan (LLNL) or Frontier (Oakridge National Laboratory). The code validation will be performed comprehensively based on comparisons to the currently available ab-initio (TD)DFT codes such as Quantum-Espresso, VASP, Octopus, Siesta, as well as comparison with experiments taking advantage of the ultrafast experimental capability at Stanford Linear Accelerator National Laboratory complemented by expert theoreticians at The Molecular Foundry.
The work was performed under the auspices of the U.S. Department of Energy by LLNL under contract DE-AC52-07NA27344 and was supported by the U.S. Department of Energy, Office of Science, Materials Sciences and Engineering Division, Computational Materials Science Program.

Robust fault detection and clustering in semiconductor manufacturing processes
LOH Woong-Kee (Gachon University)

The semiconductor manufacturing consists of a number of processes, and even a small fault occurring at any stage can damage the overall product quality. Fast and accurate detection of such faults is essential to maintain high manufacturing yields. In this talk, we present algorithms for fault detection and clustering in semiconductor manufacturing processes. The fault detection algorithm is a modification of the discord detection algorithm called HOT SAX, which adopted the SAX representation of time-series for efficient storage and computation. The clustering algorithm can be used to find the causes of faults by grouping the fault detection results. We evaluate our algorithms through experiments using the time-series data obtained from real-world semiconductor plasma etching processes. As a result, our fault detection algorithm achieved 100% accuracy without any false positive or false negative. Our clustering algorithm formed good clusters of process runs having similar sources of faults.


General relativistic radiation magnetohydrodynamics simulations of black hole accretion flows based on solving the radiative transfer equation 

ASAHINA Yuta (University of Tsukuba) 

An accretion disk is formed around a compact object such as a black hole (BH) when rotating gas accretes onto it. In order for the gas to accrete, angular momentum must be transported, and magnetic fields play an important role. We also need to consider general relativistic effects in order to solve the structure near a black hole. In addition, radiation effects cannot be ignored for very bright objects such as ultra-luminous X-ray sources. Therefore, general relativistic radiation magnetohydrodynamics (GR-RMHD) simulations are essential to study the structure near a BH. However, most of the simulations solve the radiation transport approximatively to reduce the computational cost. Hence the accuracy of the calculation in the optically thin region is reduced. Therefore, we have developed the GR-RMHD code based on solving the frequency-integrated time-dependent radiation transfer equation. In this talk, we will present the results of the BH accretion flow simulations using this code and compare them with the approximate method. Although the computational cost is high, we show that the code can accurately solve the structure of the radiation field in the optically thin region such as near the rotation axis.


Bioinformatics in the 21st century: populations, viruses and proteins for a better future
KOPELMAN Naama (Holon Institute of Technology)

The history of genetics is filled with exciting breakthroughs, among which is the emergence of Next-Generation Sequencing (NGS) technologies. The increased throughput of the sequencing data has enabled thorough investigations of the genetic variation in various species and populations including in applications such as single-cell genomics. In parallel, advances were made at the theoretical and computational level, supporting large-scale genomic analyses. In this talk I will present my research on three case studies of genetic variation analyses: 1) Jewish populations’ history, and historical questions in light of population diversity; 2) Dynamics of SARS-CoV-2 spread in Israel in spring 2020, and epidemiological reconstruction of the history of this virus; 3) Design of improved proteins for the food and beverage industry – and an improved future for us all.