Simulations in the era of exascale computing

KSK: The computational modelling of diseases will lead to transformational discoveries and therapies for diseases that are currently considered incurable, such as metastatic cancer, Alzheimer disease and autoimmune diseases. Important advancements have been made in characterization techniques, advanced animal models, early detection and therapies to address the symptoms of these health conditions. Yet the fundamental mechanisms underlying these diseases, some of which manifest at the atomic, nano- and microscale, are currently unknown or inadequately known, and need to be understood to develop cures.

The in silico investigations of diseases will be central to enabling discoveries that will lead to therapies and to eradicating terminal diseases. The investigation of these diseases will require high-resolution multiscale models of whole organs and tissues, from the atomic scale to the macroscale, which will serve as computational ‘testbeds’. The next generation of supercomputers should be able to achieve this goal in the coming decade. The trial-and-error methods practised in medicine for many conditions will then be applied to computational models rather than patients. Advances in ML and machine vision should enable computational efficiencies in the development and validation of models. For example, they would accelerate the development of accurate force fields for atomistic simulations, which would narrow down the potential protein–protein and protein–mineral interactions to take into consideration and enable the identification of potential drug molecules to test with the computational testbeds, while ensuring accurate modelling of cellular and tissue morphology, cell migration, tissue formation and tumour development.

CSC: In the next 5–10 years, I envision that thanks to exascale and post-exascale computers the plasma–material interaction physics will be understood, with all the most important multiple phenomena simulated together, including the alpha particles and their helium ash particles (generated from the fusion burn), material-sputtered impurity particles and their migration in the plasma, micro-turbulence physics, large-scale fluid-type motions and instabilities, and the behaviour of materials at the microscale.

VLD: As was mentioned by my colleagues above, a currently insurmountable problem is fully connecting the atomic and macroscopic scales in simulation. We will continue to make atomic-scale modelling much faster, but even with the fastest ML tools we are unlikely to reach the length scale of centimetres and the timescale of seconds and hours on which ‘real’ experiments often occur. These are two (currently) intractable problems. Regarding length scale, in the future, we might combine atomistic simulations with larger-scale approaches from materials science and engineering, aiming to develop unified models with adaptive resolution — fine-grained where needed, but only there. Regarding timescale, I expect that there will be a need for using advanced sampling techniques much more routinely, and maybe we will come up with entirely new ideas.

A specific issue with ML models is that the vast majority is trained for specific problems: a new domain of application requires at least the extension and re-training of an existing model, and sometimes the development of a whole new training dataset. Can we construct large ‘general chemistry’ models that are applicable to multiple scientific questions (and across the periodic table) all at once?

VVS: A very challenging problem is how to deal with materials that cannot be simulated accurately with DFT, such as materials where strong electron correlations are present. This is typically the case for materials that contain transition-metal compounds with complicated spin states or rare-earth compounds with partially filled f-electron bands. For those systems, current exchange–correlation functionals fail, and one should resort to much more expensive Green’s function or direct wavefunction-based methods, which have very bad scaling behaviour with increasing number of electrons11. This intractable scaling behaviour makes it impossible to apply these methods to realistic materials of reasonable size. ML methods may be important to enable further progress, but I also expect fundamentally new elements taken from other fields, like tensor networks inspired from quantum information theory, to play a role in the future. In any case, to progress in this field, various communities will have to cooperate, in particular quantum chemists, many-body physicists and computer scientists.

CMW: The use of DFT to predict new materials has exploded in the past few years. Only a few years ago, the DFT prediction of a novel, stable material warranted publication in a very high-impact journal. Now it’s quite routine, and papers are regularly published that predict tens or hundreds of such materials. This leads to the intriguing question: can we actually find all possible stable inorganic materials? This question would have seemed ludicrous a short time ago, but now we can actually debate the idea. What is the total number of stable inorganic compounds, and what fraction of these have we experimentally discovered? How far are we from this ‘finish line’? In just the past few years, computational predictions have blossomed, and there are DFT databases that have several times more predicted stable compounds than the total number of experimentally known ones.

Synthesizing even a fraction of these materials and achieving rationally designed, computationally guided synthesis will be a great challenge in the coming years. Pushing the science of synthesis forward is currently an active area of research, and it seems that future developments will push towards predictive, synthetic ‘recipes’ for producing novel materials. Another grand challenge involves autonomous materials discovery and development via a combination of human-out-of-the-loop experiment, computation and ML/AI. Computational methods to automate experimental characterization, accelerate the prediction of new materials and train AI approaches will play a key role in the development of this nascent field.

More advanced property prediction is also advancing quite quickly. Calculation of properties that are beyond simple DFT total energy calculations are already forming valuable datasets and are likely to form the basis of future datasets. For instance, my group is actively working in constructing high-throughput datasets that move beyond the confines of T = 0 K energetics by calculating phonon properties, both in harmonic and anharmonic forms. Harmonic enables the computation of (vibrational) free energies, and anharmonic of phonon-scattering processes and higher-order effects, such as thermal conductivity and temperature-dependent phonon renormalization.

Another means to improve these databases is to incorporate more accurate exchange-correlation functionals (such as meta generalized gradient approximation functionals) in the computational workflows. One also can imagine that AI-based approaches might learn new functionals that are physically more accurate than our best theoretical constructs. Work has already begun in this direction, but the next 5 years will be telling. Have we reached the era of AI-informed DFT or, more generally, AI-informed materials science?

Finally, all these data-driven advances will happen against the backdrop of the use of computation in high-fidelity, physics- and chemistry-informed understanding of materials and processes. These kinds of uses of computation have been present since the advent of these techniques, and have produced innumerable advances in understanding. This work will also progress alongside the data-driven work, and will probably produce new insights and discoveries and generate new understanding that is difficult to forecast.

留言 (0)

沒有登入
gif