Data-driven use cases towards Exascale

One of the primary goals in CoE RAISE is the development and expansion of AI methods along representative use cases from research and industry, which have a strong focus on data-driven technologies, i.e., analyzing data-rich descriptions of physical phenomena. Example use cases vary widely and range from fundamental physics and remote sensing to 3D printing and acoustics.

 

Use case 1: Collision event reconstruction at the CERN Large Hadron Collider

The Large Hadron Collider (LHC), the world’s largest machine, collides protons at close to the speed of light, creating conditions similar to those just tiny fractions of a second after the Big Bang. Large detectors, the size of apartment buildings, collect data from these collisions which can then be used by physicists to do their research.

proton-proton-collision.jpg

Figure 1: A proton-proton collision as recorded by the CMS experiment.

The LHC is being upgraded, which will increase the intensity of the proton beams, or the luminosity, in order to produce more collisions and therefore also more data for the physicists to analyze. This increase in data production motivates efforts to increase the speed and efficiency with which data is processed and analyzed. This is one of the major challenges facing CERN that has to be solved by the time the HL-LHC starts operation.

LHC_HL-LHC-plan.jpg

Figure 2: LHC/HL-LHC Plan. After the Long Shutdown 3 (LS3), where the HL-LHC upgrade will take place,
the production rate of data will be greatly increased.

One of the many different approaches that are being investigated in order to tackle this challenge is replacing traditional algorithms with faster, parallelizable AI versions which can be accelerated by hardware such as GPUs or FPGAs. The aim is to increase the processing speed, and thereby be able to process and analyze more data, while also keeping or increasing the accuracy of the algorithms.

One such traditional algorithm that could potentially be replaced by an AI-based version is the so-called Particle-Flow Reconstruction algorithm. Its job is to process signals from different subdetectors, such as the tracker system and the calorimeters, and combine the information in those in order to construct higher-level physics objects. In other words, it takes detector signals as input and gives particles and their corresponding momentum, energy, direction of travel etc. as output. Our job in CoE RAISE is not only to contribute to the continuous development of a Machine Learned Particle-Flow (MLPF) algorithm, but also to leverage HPC resources for efficient training and hyperparameter optimization.

In addition, the optimization of data input to AI models when performing large-scale distributed training and the porting of traditional algorithms to GPU hardware is of high importance when moving towards Exascale. The HPC Centre of Riga Technical University, a member of the RAISE project, is collaborating with CERN to investigate how to improve the clustering algorithm performance for the CMS high granularity calorimeter (HGCAL), focusing on the development of efficient and portable parallelization solutions for heterogeneous HPC hardware environments.

Use case 2: Seismic imaging with remote sensing for energy applications

 

From powering vehicles we drive to heating up our houses, we are always in need of energy sources. Besides the traditional hydro-carbon resources, nowadays there is an increased interest in geothermal heating and wind-farming. In addition, we could also use the extraction process in reverse: storing CO2 or hydrogen into subsurface reservoirs. For such operations, knowing what is located below the Earth’s surface is a crucial asset to such operations. The main methodology used to cover large areas extending deep into the Earth’s subsurface – with decent resolution – is by using sound waves, so-called seismic imaging. Here in CoE RAISE, we have scientists that employ seismic imaging techniques to reveal the underground structure of the Earth, but also experts that know the exact condition at the surface from satellite imaging. In this project we will combine both methodologies to extract the maximum information from the Earth in order to facilitate these exploration and monitoring methods. 

3D-seismic-imaging.jpg

Figure 3: 3D seismic imaging volume, showing the subsurface layering structures.

Satellite-imaging-method.jpg

Figure 4: Satellite imaging method to scan the Earth’s surface.

Earths-surface.jpg

Figure 5: Satellite image showing the different properties of the Earth’s surface.

Use case 3: Defect-free Metal Additive Manufacturing

 

Within the past 10 years, we have seen the meteoric rise of 3D printing technology (also called additive manufacturing), which is now being used even in fields that impose the most stringent quality requirements, e.g., aerospace engineering. But quality control for the parts that are produced remains challenging. In CoE RAISE, we study in particular industrial 3D printing to build metallic parts, where metal powder is locally melted together by a laser, to form a solid object layer by layer, a process known as selective laser melting (SLM). Within this task, we focus on developing an in-process monitoring system that allows to perform quality control during the manufacturing process.

Metal-3D-printing-experiment.jpg

Figure 6: Metal 3D printing experiment.

At each point in time, the high-powered laser moves over a spot in the powder bed, which melts quickly. Occasionally, metal vapor bubbles may cause spatters, molten pieces of metal which get ejected from the melt pool. Such observations, and also more subtle phenomena and interactions, can be correlated with later quality issues. 

molten-metal.jpg

Figure 7: The molten metal radiates brightly.

Metal-vapor-bubbles.jpg

Figure 8: Metal vapor bubbles causing spatters, molten pieces of metal get ejected from the melt pool.

The events in the melt pool are quite fast. Therefore, we have instrumented Flanders Make’s experimental platform based on an industrial 3D printer with an additional high-speed camera system, which can record the events in the melt pool at 30,000 frames per second. The camera can be mounted to record overview images of the powder bed, or on a half-mirror in the optical path of the laser to provide close-ups centered around the laser beam (as in the illustration above).

The printing of even a small (cm-scale) object yields terabytes of raw data. The hypothesis space is also large because later events may influence and even undo the results of previous events. Therefore, the analysis and predictive modeling benefits greatly from HPC.

Use case 4: Sound Engineering

 

The shape of your ears is as individual as your fingerprints and alters every sound you hear. Characterizing this effect, the head-related transfer function (HRTF), for a given person allows computers to present an accurate virtual sound environment.  It has been used, for example, to help visually-impaired people navigate unknown environments. At the moment, however, collecting an individual's HRTF requires cumbersome direct measurement in a specialized facility.

At the Simulation and Data Lab ACUTE, a member of the RAISE project, we are investigating how the geometry of your ears affects your hearing.  To do this, we collect a database of ear shapes and their corresponding acoustic properties. This database is then used to train an AI which can predict your HRTF from a 3D scan of your head.  This scan can be produced anywhere with only a handheld scanner, which allows HRTF-based technologies to leave the lab and provide benefits to the wider population.

Ear-scans-and-models.jpg

Figure 9: Test setup in the anechoic chamber (left and top). Ear scans and models (bottom middle and bottom right).

Collaboration helps everyone taking part

 

The use cases in WP4 can, at a first glance, seem very different from one another, but in fact have much in common. They face the same challenges regarding the handling and analysis of very large datasets and how to best extract useful information from the data. These challenges will become even more important to tackle efficiently when we operate at Exascale. By collaborating in CoE RAISE, partners from different use cases share their experiences with each other and thereby gain new perspectives on the problems they are trying to solve. This gives rise to a collision of different viewpoints of the same or similar problems which in turn has the potential to bring new and innovative ideas to life.

Links

Overview of all CoE RAISE use cases

Social Media Chanels