ll

OPEN ACCESS

Editorial
Accelerating scientific discoveries through data-driven
innovations
Developing artificial intelligence (AI) and machine learning (ML) While the development of novel AI/ML techniques is rapidly
methods that can accelerate scientific discoveries and advance transforming scientific research, there is no doubt that such
science has become one of the important research directions for transformation brought forth by rapid advances in AI/ML would
the AI/ML research community. It has been gaining increasing not have been possible without advances in modern HPC. Moti-
attention from researchers in diverse scientific areas, including vated by the fact that the semiconductor industry is getting
biomedical science, materials science, climate science, physics, closer to the limit of chip miniaturization, Weidner et al. explore
chemistry, and many others. Data-driven AI/ML innovations to quantum computing as an alternative paradigm for simulating
enable reliable predictions and optimal decision making for sci- complex systems. Based on gene regulatory networks modeled
entific discoveries face several critical challenges, among which by Boolean logic, the authors demonstrate how a quantum algo-
are high system complexity, large search space, incomplete rithm using qubits can effectively analyze the network dynamics
knowledge, and small data, all of which demand novel strategies despite the exponential growth of network states and the simu-
to effectively address them. Meeting these challenges and lation complexity.
thereby accelerating scientific discoveries and industrial innova- An important application area for AI/ML that is receiving
tions, calls for research that can take full advantage of the latest growing attention in science and engineering is the utilization
advances in AI/ML to integrate data-driven techniques with sci- of advanced AI/ML techniques for real-time monitoring, safe
entific knowledge and is able to execute them in modern high- and efficient management, and autonomous control of complex
performance computing (HPC) environments at scale. This devices, equipment, or facilities. In the paper entitled ‘‘Long-
Patterns special collection ‘‘Accelerating scientific discoveries sequence voltage series forecasting for internal short circuit
through data-driven innovations’’ features articles that show- early detection of lithium-ion batteries,’’ Cui et al. showcase
case the promising roles of AI/ML and data-driven modeling in the use of AI/ML for early detection of internal short circuits
accelerating scientific discoveries and may inspire the next (ISCs), a common safety concern for lithium-ion batteries. Using
wave of data-driven innovations in various scientific domains. a deep learning scheme based on an encoder-decoder architec-
AI/ML models are becoming increasingly popular in materials ture, the method is capable of predicting long-sequence voltage
science research as they enable computationally efficient pre- and power series where the evolution inconsistency between
diction of material properties and thereby can facilitate the dis- normal and ISC battery voltage can be used for accurate ISC
covery of novel materials with desirable properties. Effective detection. Similarly, one may envision the use of AI/ML for
property prediction using AI/ML models requires accurate struc- continuously monitoring the health, safety, and efficiency of
tural information of the material, which is unavailable for new complex systems—such as wind turbines or power grids—as
materials. However, ab initio prediction of equilibrium structures well as the autonomous control of experimental facilities—such
is computationally expensive, creating a significant bottleneck as complex beamlines and associated detectors—for scientific
in materials discovery pipelines. In the paper ‘‘Strain data experiments.
augmentation enables machine learning of inorganic crystal In recent years, data-driven methods for molecular optimiza-
geometry optimization,’’ Dinic et al. address this challenge by tion have received increasing attention—especially for applica-
training an ML model that can predict the impact of local and tions such as drug discovery and materials design—because
global strains and enhance energy predictions for structural de- of their potential to significantly accelerate the discovery of novel
formations. The trained model allows efficient ML-based geom- molecules with desired properties. Devising effective methods
etry optimization that leads to improved prediction of formation for multi-objective optimization is critical in molecular design,
energy for inorganic crystals. as one typically needs design molecules that satisfy multiple—
A common challenge in leveraging AI/ML in scientific applica- and often competing—design criteria. In the review article enti-
tions is the scarcity of relevant data that can be used to robustly tled ‘‘Computer-aided multi-objective optimization in small
train accurate data-driven models. One potential approach for molecule discovery,’’ Fromer and Coley focus on Pareto optimi-
alleviating this issue is domain adaptation (DA), which refers to zation algorithms for multi-objective molecular design. Although
a technique that aims to improve the performance of a model it is commonplace, mainly for simplicity, to combine multiple
in a ‘‘target’’ domain of interest that is different from the ‘‘source’’ properties via scalarization to define a single objective function,
domain in which the model was originally trained. Typically, DA Pareto optimization methods have important advantages as they
techniques can be helpful for training an ML model in a data- obviate the need to pre-determine the relative importance be-
poor ‘‘target’’ domain by leveraging the data in relevant ‘‘source’’ tween different properties and as they can reveal the trade-offs
domain(s) that may be relatively data rich. The work by Wu et al. between competing properties. In addition to providing a
proposes an optimal transport method, called MROT, that en- comprehensive review of both pool-based and de novo genera-
ables unsupervised and semi-supervised DA for molecular tive approaches for multi-objective molecular design, the au-
regression tasks to facilitate the exploration of new molecules thors also discuss open problems and opportunities for future
in small data regimes. research.

Patterns 4, November 10, 2023 ª 2023 The Author(s). 1

The review paper by Qian et al. entitled ‘‘Knowledge- This work shows that by optimizing the operational policy (i.e.,
driven learning, optimization, and experimental design under the screening policy in this study), one can substantially
uncertainty for materials discovery’’ discusses several salient accelerate screening virtually without any degradation in the
challenges in applying data-driven methods to real-world com- screening accuracy.
plex systems and presents recent advances in the field for ad- Although the papers included in this Patterns special collec-
dressing them. The central issue is the complexity of real-world tion showcase some of the fascinating ways that data-driven in-
systems for which existing data may be insufficient for accurate novations can accelerate scientific discoveries, they are by no
system modeling or identification. Moreover, acquisition of means comprehensive. We expect that we will see continued
ample data may be practically difficult or formidably expensive, significant data-driven discoveries and rapid advances across
as a result of which existing ML methods may fail to produce reli- diverse scientific domains in the upcoming years through
able and generalizable predictions. Qian et al. show how uncer- cross-cutting research at the interactions of AI/ML, mathe-
tainty-aware and objective-driven Bayesian signal processing matics, and computing.
and machine learning schemes can enable effective optimization
and experimental design for complex uncertain systems.
DECLARATION OF INTERESTS
Furthermore, this review article showcases how knowledge-
driven and physics-informed approaches can be critical for B.-J.Y. is a member of the Patterns journal advisory board.
effective learning in data-limited scientific domains.
High-performance computing has been playing central roles in
various scientific domains, and its importance has been growing Francis J. Alexander
even further as a result of the increasing adoption of AI/ML Computing, Environment and Life Sciences, Argonne National
models in scientific inquiries. Executing large-scale computa- Laboratory
tional campaigns requires decision making to efficiently allocate Meifeng Lin
the computational resources to ensure that we obtain the Computational Science Initiative, Brookhaven National Laboratory
desired outcomes while minimizing the execution time or Xiaoning Qian
computational cost that is incurred. Traditionally, such decision Department of Electrical and Computer Engineering, Texas A&M
University
making has relied on the intuition and experience of domain ex-
Computational Science Initiative, Brookhaven National Laboratory
perts, which may yield reasonably good but suboptimal Byung-Jun Yoon
results. In the paper ‘‘Optimal decision-making in high- Department of Electrical and Computer Engineering, Texas A&M
throughput virtual screening pipelines,’’ Woo et al. tackle this University
optimal computational campaign (OCC) design problem in the Computational Science Initiative, Brookhaven National Laboratory
context of high-throughput virtual screening (HTVS) campaigns. [Link]

2 Patterns 4, November 10, 2023