A common problem in numerical optimization and sampling is the detection of relevant states. These could be, for instance, the local minima on a rugged parameter surface or the transition state of a chemical reaction. For most cases, an exhaustive search for the optimal solution is intractable. Here, we focus on parallel sampling and optimization strategies relying on multiple replicas, most prominently, adaptive methods where all simulated replicas use the same propagator and sample the same underlying surface. In these methods, replica intercommunication is used to provide a global assessment as to which replicas are most interesting. This implies, in general, periodic data mining steps across replicas. Furthermore, in order to extract and utilize the gained information in post-processing, data must often be stored, which poses stringent data management and analysis challenges in particular for high-dimensional cases. The minisymposium wishes to discuss the following questions: What are meaningful and easily generalizable tools, strategies, and algorithms to guide the sampling/exploration? How can we maintain scalability and load balance? What types of post-processing algorithms can be applied to the generated data, and are those scalable to provide on-the-fly solutions to direct the exploration?

Selected PASC projects will present the scientific challenge they aim to solve by using high-end supercomputers, and in particular the computational approach adopted. The topics include astrophysics (smooth particle hydrodynamics), numerical weather and climate (stencils and grids), linear algebra for electronic structure (sparse matrix and tensor operations), and biomedical applications (fluid-structure interaction, machine learning).

The focus will be on aspects concerning (i) capability computing: how to scale to several hundreds/thousands of compute nodes, including the use of communication optimal algorithms and asynchronous communication; (ii) performance portability: how to address the growing diversity in hardware on a compute node, including generic software design, and auto-tuning; and (iii) co-design in these projects: how to engage with vendors to optimally exploit current hardware, and to provide feedback that has or will influence next-generation hardware.

Topics include: side by side comparisons of multi-core, many core, and GPU compute nodes; optimization techniques for flops, memory bandwidth, or network performance; JIT compilation of machine specific kernels; programming approaches such as the use of domain specific languages (DSLs), remote memory access (RMA), task based programming.

Discrete-time, infinite-horizon, general equilibrium models are routinely used in macroeconomics and in public finance for exploring the quantitative features of model economies and for counterfactual policy analysis. One important question concerns the importance of household heterogeneity for the amplification and propagation of macroeconomic shocks.

In this session we bring together leading young researchers in the field to present alternative approaches to the computation of equilibria in dynamic stochastic models with heterogeneous agents and/or with financial frictions. Three of the papers directly propose new methods for the solution of models with a continuum of ex post heterogeneous agents.

The power of artificial neural nets at executing challenging tasks learned from data is very attractive to fields of physical science and especially to High Energy Physics. In recent years, there have been a significant number of articles reporting promising results with applying deep learning to HEP challenges. By virtue of the very large number of parameters of artificial neural nets, with deep and wide architecture, trained with stochastic gradient descent, it is mandatory to process a lot of representative data in order to obtain accurate models. Training of such models requires a tremendous amount of computing, and commonly takes days, if not weeks, to converge. GP-GPU technology has enabled a lot of this computation, thanks to the high level of parallelisation of the formalism of artificial neural net, but more however can be gained in parallelised calculation of the stochastic gradient descent. Supercomputing facilities are particularly suited for distributed training of deep neural nets, thanks to their large computation power and excellent connectivity. This minisymposium will address the current state-of-the-art and present performances in training models for high energy physics, with a particular view to software availability and to foster further utilization of supercomputers.

Achieving hardware and energy efficiency is important for current large-scale numerical simulations and will be a key component in the exascale era. In a world of heterogeneous, highly parallel computer architectures with deep memory hierarchies, complex application scenarios, and a broad spectrum of algorithms, a thorough analysis and understanding of the complex interaction of software, data structures, algorithms, and hardware features, a.k.a. performance engineering, is required for implementing codes that allow for portable performance on the computer generations to come. The minisymposium addresses a broad range of topics in performance engineering for modern HPC architectures, ranging from recent advances in performance models and tools supporting a "white-box" performance engineering approach to application performance tuning cases studies and "black-box" solutions. The presentations will point out the potentials and limitations of performance engineering activities and demonstrate the wide spectrum of performance models used in the performance engineering, including simple performance expectations, automatic model parameter selections, and analytic models.

Modern electronic-structure methods provide parameter-free simulations of properties across the whole spectrum of Physics, Chemistry, and Materials Science. They have become a bedrock of advanced analytical methods and are used systematically to interpret advanced experiments and complex interactions, with ever growing perspectives for more "realistic" systems, including defects, thermal and external fields, and transient phenomena. This minisymposium explores the next generation of electronic-structure software, which will lead users to exascale supercomputers, through highly efficient and highly parallel algorithms. We will showcase recent advances in ground- and excited-state calculations, spectroscopic quantities and transport, and adaptive methods which exploit different algorithms for different systems such as periodic/localized or many/few electrons per atom.

The increasingly large amounts of data being produced by weather and climate simulations and earth system observations is sometimes characterised as a deluge. This deluge of data is both a challenge and an opportunity. The main opportunities are to make use of this wealth of data to 1) improve knowledge by extracting additional knowledge from the data and 2) to improve the quality of the models themselves by analysing the accuracy, or lack thereof, of the resultant simulation data. An example of the former case is improved prediction of large scale phenomena such as El Nino. An example of the latter is the improvement of a Physics parameterisation scheme through detailed analysis of the errors in a large number of datasets.

One way to realise these opportunities is to use machine learning approaches. As machine learning in weather and climate is a relatively new topic this minisymposium introduces the audience to how machine learning could be used in weather and climate and outlines its implications in terms of computing costs. To ground the ideas in concrete examples it also illustrates the use of machine learning in the weather and climate domain with practical examples.

Wall resolved LES is prohibitively expensive, and a first principles based wall modeled LES, that is free of tunable parameters, is far from becoming an alternative to RANS as a predictive tool for design. On the spatial discretization front, most higher-order and spectral methods have had many of their successes in the realm of low to moderate Reynolds number flows, and on relatively less complex geometries. The bulk of the complex flow high Re simulations still employ nominally second-order accurate schemes in unstructured mesh finite volume flow solvers, and finite-element based solvers with modest polynomial order. Even in these fairly mature solvers, evidence suggests that merely running larger cases with an increased number of grid points and on larger computational domains does not always guarantee a better solution (in the LES sense). There is, therefore, a need for a two-level effort whereby benchmark higher-order simulations can serve to inform sub-grid models to improve the predictive capability of existing mature flow solvers. We hope to motivate a discussion along these lines with examples of results of higher-order simulations, as well as their role in assessing and improving the predictive capabilities of sub-grid models with more conventional flow solvers.

A common problem in numerical optimization and sampling is the detection of relevant states. These could be, for instance, the local minima on a rugged parameter surface or the transition state of a chemical reaction. For most cases, an exhaustive search for the optimal solution is intractable. Here, we focus on parallel sampling and optimization strategies relying on multiple replicas, most prominently, adaptive methods where all simulated replicas use the same propagator and sample the same underlying surface. In these methods, replica intercommunication is used to provide a global assessment as to which replicas are most interesting. This implies, in general, periodic data mining steps across replicas. Furthermore, in order to extract and utilize the gained information in post-processing, data must often be stored, which poses stringent data management and analysis challenges in particular for high-dimensional cases. The minisymposium wishes to discuss the following questions: What are meaningful and easily generalizable tools, strategies, and algorithms to guide the sampling/exploration? How can we maintain scalability and load balance? What types of post-processing algorithms can be applied to the generated data, and are those scalable to provide on-the-fly solutions to direct the exploration?

Numerical weather prediction and climate models are large and complex software applications that need to run efficiently on today's and future massively parallel computer systems. The rapid change in these computing architectures and the increase in diversity are seriously affecting the ability to retain a single source code that runs efficiently in different architectures. Several weather models have successfully adapted their codebases to many-core and heterogeneous architectures like GPUs and Xeon Phi using a combination of multiple traditional programming models for parallel architectures like OpenMP, OpenACC and MPI. However porting existing large community codes to multiple architectures is a daunting task and leads to codes that are more complex and difficult to maintain. As a result in the past years numerous new technologies and approaches are emerging in order to provide new programming models, like domain-specific languages (DSLs) or source-to-source translation tools that can increase the productivity of development in weather codes while providing a high degree of performance portability. In this minisymposium we propose a discussion with some of the most prominent novel approaches where the new advances in programming models used for heterogeneous architectures in weather and climate models will be presented.

Many important economic phenomena relate to the notion of risk. Economic actors not only make decisions as a function of their current situation but also depending on their expectation of future developments. Because economic systems are not deterministic, future economic events are usually treated as stochastic phenomena. The general form of the problem at hand is to determine how the probabilistic distribution of future economic events influences current decisions. The wider the distribution, the more risk in today's decisions. The papers in this session present different computation challenges involved in attempting to describe the effect of risk on economic decisions.

The dawn of Web 2.0 applications, smartphones/tablets and the omnipresent yet invisible cloud computing have dramatically changed our perception of the IT landscape in the last decade. At the same time these technologies delivered an abundance of new tools that are more suited to the needs of domain scientists: GitHub/GitLab with their inherent support for agile programming, user space package managers for easy software installations even when facing complex dependencies, and web portals to HPC systems or private clouds so that no prior knowledge is needed to use state-of-the-art compute resources. This minisymposium will showcase all these tools and how they are used in real scientific workflows. The shown best practices are easy to reproduce since they are all based on freely available software packages, so that the audience can use the presentations both as an inspiration but also as a kickstart for their own better science.

The Large Hadron Collider at CERN is smashing high density bunches of protons near the speed of light at a frequency of 40 MHz. Most of the thousands of particles emitted at each bunch crossing are measured and collected with building-sized detectors consisting of multiple sub-detectors each serving its own purpose. The simulation of the signal created by a particle interacting with such a detector is typically done with very detailed simulations and needs to be stepped infinitesimally over meters of material. This simulation is as much computing intensive as the geometry is complex. In current and future detector design, the fine-grained simulation of such a detector is taking a great part of the full computing budget of experiments and poses a computing challenge. While a great deal of effort is being made to parallelise such software, one possible avenue to reduce the computational requirements is with generative models from the field of deep learning. Such generative models have seen success in conditionally generating images and video of various types. We present how such models are built and trained, and how well they can capture the physics of particle interaction and help generate realistic samples for high energy physics analysis.

In the past couple years a large number of fintech and more recently insurtech startups have been founded and are challenging established players in the financial services sector. At Baloise – a Swiss company providing insurance services in Switzerland, Belgium, Germany and Luxembourg as well as banking services in Switzerland – we view startups as potential partners on our digital transformation journey rather than competition. This minisymposium aims to demonstrate what problems companies such as Baloise face in terms of digitizing their business and making use of their large amounts of data. To do so, the four sessions will cover the innovation framework Baloise employs in order to rapidly test prototypes, a presentation by Brainalyzed, a startup which aims to optimize and automatize investment decisions at Baloise using AI, a presentation about the challenges arising in the context of data warehouses and legacy systems, and a panel discussion with all the speakers.

Machine Learning is right now a booming field of computer science which finds applications in the development of computer-human interfaces, in the analysis of medical data of huge populations, in the maintenance of cars, planes and elevators, and self-driving cars, to mention a few. During the last twenty years the field has gone through a development and refinement which has been spectacular. For some reason the use of the technology in pure science has been lagging behind; however, we are now starting to see the use of machine learning in the field of quantum chemistry. Here, the approach will enhance the performance of standard quantum chemical calculations – improving convergence, could serve as a tool for post-analysis of huge sets of ab initio results, or could simply replace computationally expensive procedures. Machine learning offers practical alternatives where standard quantum chemical simulations would be prohibitive. During the last few years a small number of quantum chemistry groups have explored the potential of machine learning – the results have been extraordinary and spectacular. Here in this minisymposium we would like to inspire by presenting four different applications in which machine learning is fundamental to success.

Difficult combinatorial problems permeate virtually every area of the sciences, business, and government and many of these problems can be cast as mixed-integer programs (MIPs). A MIP is a mathematical definition of a problem that is comprised of a set of constraints and an objective function. In general, MIPs are NP-hard and require exponential amounts of computation time in the worst case. However, search strategies, such as branch-and-bound, branch-and-cut, and cut-and-solve have evolved to provide optimal solutions for many instances.

Although there has been great progress in this field and computational power has dramatically increased over the years, many important MIPs remain intractable and the use of massive parallelization appears to be a promising means to address this great need. However, many challenges lie ahead. This minisymposium will elucidate some of these challenges, while highlighting progress in this field. It includes a round table discussion with Michael Chan, Sharlee Climer, Daniel Jacobson, Sarah Powers, and Daniel Rehfeldt, and is open to conference participants. The goal of the discussions will be to explore and integrate high-performance expertise with domain-specific insights with an aim to identify strategies that may resolve these pressing challenges.