The Rising Complexity in Science
Advancing research and innovation is fundamental to modern civilization. At its practical best, scientific research has the power to create prosperity, alleviate suffering, and raise living standards and wellbeing across the globe. And the impact of science at the more abstract intellectual level is just as significant, in that it provides the tools for us to satisfy our innate curiosity and our need better to understand the world we inhabit. In the absence of scientific progress, as a species, we would be severely diminished.
Over the past century we have made incredible progress in all areas touched by science. But while the number of people engaged in R&D today is higher than ever before, there is rising concern from some quarters that the pace at which important discoveries are being made may be slowing . The complexity of science is rising rapidly. And at the same time, R&D is becoming increasingly expensive , both in terms of developmental timescales and human capital, and of the tools needed to support groundbreaking advances. The confluence of increasing R&D costs and slowing pace of innovation is well known in the context of pharmaceutical R&D efficiency, and has been dubbed Eroom’s Law : a literal reversal of Moore’s famous law on transistors. While this bleak prospect is not (yet) representative of the broader scientific landscape, it is essential to explore what mitigating strategies must be put in place today to ensure that stagnation, cost, and most importantly of all – complexity, do not stifle future innovation and scientific research. Among such Eroom-breaking strategies, computation stands out as being of particular importance. By computation I mean the general ability to use computers and numerical models to mimic real systems, and explore them in silico before making final design decisions or undertaking costly realisations in the laboratory.
The importance of computation in Physics has been acknowledged from very early on, and has been a prominent part of the research environment for at least the last 50 years. But the increase in computational power over the past couple of decades, alongside the lowering of cost barriers, has really started to make many promising ideas realisable in practice for the first time only relatively recently. Computation is, arguably, our single most important exploration tool in other areas of science too; from the design of fusion experiments to semiconductor process design, chemistry, and advanced materials modelling, the vast majority of research areas today would be woefully incomplete without at least some part of computational simulation.
“Among such Eroom-breaking strategies, computation stands out as being of particular importance.”
One extreme representation of computationally-aided discovery is the concept of a digital twin: an accurate virtual representation of a system, which can be probed, manipulated and perturbed at will to yield new understanding. This is a powerful paradigm, but it is worth pointing out that in the sciences it still remains unrealisable in all but the simplest systems. And more traditional modelling approaches remain saddled with significant limitations too. Quantum chemists, for example, do not have the predictive power to simulate even relatively small molecules at the correct temporal and spatial scales, and a multitude of approximations, assumptions, and ad-hoc adjustments are needed for practical applications. More generally, we remain largely unable to make effective use of long-scale quantum coherence in materials design. Or take, for example, inertial confinement fusion: despite groundbreaking recent advances , it is still not possible to simulate the entire implosion dynamics of a fuel capsule on the National Ignition Facility at the level of resolution needed to accurately model the full range of plasma instabilities. This makes a model-driven approach to the design of fusion experiments extremely difficult. In fact, the winning strategy has often been to adapt the experiments to configurations that are easier to simulate, rather than optimising the experiments themselves, effectively turning the design-optimization paradigm on its head. For as long as we lack the computational power to model complex systems at scale, human ingenuity will need to fill the gap, if at the great cost of added complexity.
Within the broader area of computation, our Eroom-breaking armamentarium consists of several elements. First and foremost, a step change in raw computational capability remains both the most immediate and the most general. This is closely linked to a continuation of Moor’s law into the future, itself a fascinating challenge. Considerable gains remain to be extracted on the hardware side via increases in both chip capacity and complexity. New chip designs suited to specific problems may also provide non-negligible boosts to bottleneck tasks. But it will be imperative for scientific research to be able to tap into this rapidly growing area and reap the full benefits. Encouragingly, science has a strong ally in the semiconductor industry, who share this vision that places “semiconductors at the centre of the next evolution for humankind” , to paraphrase Aart de Geus at a recent meeting of the President’s Council of Advisors on Science and Technology.
On the software side, new algorithms will continue to improve the efficiency and robustness of simulation suites that underly much of what we do. And, of course, there is machine learning (ML), in all its varied richness, which, while still at the early stages within scientific research, has already shown a capacity to both accelerate simulation workflows and optimization tasks, and to improve prediction accuracy.
“Researchers need to be able to access the full power of these new technologies in ways that do not presuppose deep technical expertise that is at best tangential to their scientific objectives.”
However, for these exciting developments to live up to their promise, they have to be part of an integrated solution to the growing complexity of science, rather than an addition to it. If the only way to tap into this pool of opportunity is to become an expert in ML or in algorithm implementation, progress will stall. Researchers need to be able to access the full power of these new technologies in ways that do not presuppose deep technical expertise that is at best tangential to their scientific objectives. In data science, where one seeks to extract information from large and often incomplete datasets, several companies have realised the considerable value in bridging this gap between technology and application via easy-to-use data platforms. In contrast, very little effort has been made so far in the space of computational modelling. Here, scientists and researchers are still largely left to their own devices, and there remains a substantial untapped benefit to be extracted from bridging this gap. In particular, IT systems that collect, standardise, implement and deploy algorithms, methods, ML tools and simulations – in the form in which they need to be used in practice – would be transformational. A move from a world of expensive and hard-to-access know-how, to one where technical expertise is an easily accessible commodity, is going to be essential to ensure innovation wins, and continues to win, the battle over soaring complexity.