New research brings machine‑learning‑based physics a step closer to solving real engineering challenges
A mathematics professor at The University of Manchester has developed a novel machine-learning method to detect sudden changes in fluid behaviour, improving speed and cost of identifying these instabilities and overcoming one of the major obstacles faced when using machine learning to simulate physical systems.
Computational simulations of mathematical models of fluid flow are essential for everyday applications ranging from predicting the weather to the assessment of nuclear reactor safety. The advent of this simulation capability over the past 50 year has revolutionised the development of fuel-eļ¬cient aeroplanes and sail configurations on racing yachts can now be optimised in real time, providing the marginal gains needed to win races in the Americas Cup.
Optimised aerodynamics means that modern day cyclists can ride faster, golf balls fly further and Olympic swimmers consistently set world records. Computational fluid dynamics also enables the modelling of the flow of blood in the human heart, making the provision of patient-specific surgery possible.
Scientists and engineers rely on computer-based simulations to understand, predict, and design these systems that they can’t easily test in real life. But traditional fluid‑simulation methods often require hours or even days of computation, and struggle when the flow becomes fast or highly complex.
Machine‑learning‑based simulations, once trained, can make these assessments almost instantly. Instant feedback would allow rapid design testing, real‑time adjustments, and rapid testing variation without the usual computational burden.
The findings were published in the Journal of Computational Physics.
Professor David Silvester, Professor of Applied Mathematics at The University of Manchester said: "Solving reliability issues that machine‑learning models encounter would offer major benefits for scientific research and engineering. The issue to be faced is that naive AI predictions of flows generated solely from data are highly likely to feature impossible scenarios. This is a serious concern when predicting extreme events like tornados and tsunamis.”
The study uses the stability of fluid motion as the foundation for a new method that predicts how complex systems behave. Instead of relying on costly laboratory experiments, solutions to the fundamental equations of fluid motion are generated numerically. This allows the machine-learning model to be trained on accurate, high-quality data drawn directly from physics, demonstrating that the model can accurately handle challenging simulations.
A key focus of the work is identifying bifurcation points –the moments when a smooth, steady flow (laminar flow) suddenly begins to change – similar to calm, evenly flowing river as it hits an obstruction, or splits and fluids start to mix and form eddies. Laminar flow is when a liquid behaves in a smooth and orderly way, like pouring honey, the flow is consistent and steady.
By successfully using a machine‑learning model to identify the points at which a system changes behaviour or in this case bifurcates, the study suggests that, with further refinement, machine‑learning‑based models could become a practical alternative to traditional fluid‑modelling techniques in the future.
Professor Silvester added: "This marriage of old and new approaches holds the promise of efficient computation of physically realistic fluid flows in a myriad of practical situations. The development of refined mathematical models of complex fluids is likely to be critically important if the promise of AI is to be effectively realised in the future.”
Journal
Journal of Computational Physics
Method of Research
Computational simulation/modeling
Subject of Research
Not applicable
Article Title
Machine learning for hydrodynamic stability
Reliable material databases bridge AI- and experimental-led material discovery
Materials databases lie at the heart of future data-driven discovery in energy-related fields, say researchers from Tohoku University.
image:
The evolution of materials science paradigms.
view moreCredit: Li et al.
Materials databases lie at the heart of future data-driven discovery in energy-related fields, say researchers from Tohoku University.
In a new article published in the journal Precision Chemistry, they examined how different types of databases, both computational and experimental, work together to support modern artificial intelligence (AI) tools used in materials science.
The study found that materials databases are no longer just places to store information. Instead, they play a central role in determining how well AI models perform. The way data is collected, organized, and shared - known as database architecture - can directly affect whether AI systems produce reliable and useful results.
"In a library, if books are poorly labeled, have missing pages, or are difficult to access, even the most skilled reader will struggle to find accurate information," stresses Hao Li, lead author of the paper and Distinguished Professor at Tohoku University's Advanced Institute for Materials Research (AIMR). "In the same way, AI models depend on well-structured and carefully curated data to make sound predictions."
Li and his team categorized computational databases into two main groups: those that focus on bulk material properties and those that focus on surfaces and interfaces. They also reviewed experimental databases that cover areas such as crystal structures, catalysis, energy storage, and materials characterization.
Further analysis revealed the growing importance of integrated platforms. These systems connect computational predictions with detailed experimental data, allowing scientists to test ideas, refine models, and validate results in a continuous cycle. This approach supports more efficient and reliable materials discovery.
Moreover, the researchers introduced a roadmap for combining databases, AI models, and experimental workflows. This includes the use of graph neural networks, machine learning interatomic potentials, and large language model-based AI agents to accelerate the discovery process while maintaining scientific rigor.
However, the researchers identified several challenges that must be addressed. These include the need for standardized data practices aligned with FAIR principles (Findable, Accessible, Interoperable, Reusable), better tracking of data origins, and improved reporting of negative results, which are often missing but are important for reducing bias.
"Materials databases are the foundation of trustworthy AI in science," adds Li. "If we want AI to guide discovery in a reliable way, we must first ensure that the data it learns from is complete, transparent, and well-structured. Without reliable data, AI-led discovery will itself become unreliable."
Looking ahead, the team plans to improve database quality and connectivity across fragmented data sources. They also aim to develop new AI systems that can learn from multiple types of data simultaneously and work alongside experiments and human researchers. These efforts are expected to support more dependable and efficient discovery of materials for energy, sustainability, and everyday applications.
Journal
Precision Chemistry
Article Title
Materials Databases: Foundations of Modern Digital Materials
Computational and integrated platform.
Database-to-model-to-experiment roadmap for domain models and AI Agents.
Credit
Li et al.
Li et al.
No comments:
Post a Comment