Featured
Tyler O'Neal, Staff Editor LATEST

Physics-trained ‘Digital Super Brain’ learns from supercomputers to accelerate discovery

Chalmers researchers combine HPC simulations, photonics physics, and AI to create a new generation of scientific surrogate models

For decades, the world’s most powerful supercomputers have functioned as scientific laboratories, enabling researchers to explore complex phenomena ranging from climate systems and fusion plasmas to advanced materials and photonic devices. However, even these high-performance machines face a critical bottleneck: high-fidelity simulations are computationally expensive, often requiring days or weeks of processing time to evaluate a single design space.
 
To address this challenge, researchers at Chalmers University of Technology in Sweden have developed a "digital super brain." By integrating artificial intelligence with fundamental physical laws, they have created a machine-learning framework capable of replicating complex electromagnetic simulations with only a fraction of the computational effort. This innovation marks a paradigm shift in scientific computing, as supercomputers transition from merely performing simulations to training intelligent models that can explore new designs at unprecedented speeds.

Teaching AI the laws of physics

The Chalmers team’s breakthrough stems from a simple observation: most AI systems spend enormous effort learning physical relationships that scientists already understand.
 
Traditional neural networks are often treated as black boxes, requiring vast amounts of training data before they can accurately predict physical behavior. Generating that training data often requires thousands of large-scale simulations to run on HPC systems.
 
Instead of forcing the AI to learn everything from scratch, the researchers embedded physical knowledge directly into the neural network architecture.
 
Published in Laser & Photonics Reviews, the study introduces a framework that incorporates the physics of optical resonances through quasinormal mode (QNM) theory. The model learns the resonant behavior of photonic structures while automatically respecting fundamental physical principles such as energy conservation and causality.
 
By integrating known physics into the learning process, the researchers created a model that requires significantly less training data while maintaining high predictive accuracy.
 
For computational scientists, this represents an important shift. Rather than replacing physics with AI, the framework fuses the two into a single computational system.

Supercomputers become teachers

The most intriguing aspect of the work may be its relationship with high-performance computing.
 
The AI model depends on large quantities of training data generated through sophisticated electromagnetic simulations. These simulations, which solve Maxwell’s equations across complex nanostructured devices, are precisely the type of workloads that consume substantial HPC resources.
 
In effect, the supercomputer acts as a teacher.
 
Large simulation campaigns generate enormous datasets describing how light interacts with advanced photonic structures. The AI system then compresses this knowledge into a compact surrogate model capable of reproducing the behavior of those systems almost instantly.
 
This emerging workflow is rapidly becoming one of the most important trends in computational science:
  1. Run large-scale simulations on HPC systems.
  2. Generate high-fidelity physical datasets.
  3. Train physics-informed AI models.
  4. Deploy surrogate models for rapid design exploration.
The result is a powerful form of computational knowledge compression. Months of simulation effort can be distilled into an AI model that delivers answers in seconds.

Accelerating photonics research

The researchers demonstrated their framework on photonic crystal slabs and free-form dielectric metasurfaces, structures that manipulate light at the nanoscale and play important roles in sensing, communications, imaging, and quantum technologies.
 
Designing such devices typically involves searching through enormous parameter spaces while repeatedly running computationally intensive electromagnetic simulations.
 
For advanced photonics research, the computational burden can become overwhelming.
 
A single optimization campaign may require thousands of simulations, each demanding significant CPU or GPU resources. As device complexity increases, the associated HPC requirements grow accordingly.
 
The Chalmers approach dramatically reduces that burden.
 
Because the neural network understands the underlying physics, it can accurately predict device performance with far fewer training examples than conventional machine-learning models. This translates directly into lower computational costs and faster development cycles.

The rise of physics-informed AI

The research reflects a broader movement across scientific computing.
 
For years, the dominant trend in AI has been scaling: larger models, larger datasets, and larger computational budgets. Scientific computing is beginning to explore a different path.
 
Instead of relying solely on more data, researchers are increasingly embedding scientific knowledge directly into machine-learning systems.
 
The advantages are substantial:
  • Improved accuracy
  • Better interpretability
  • Reduced training requirements
  • Stronger adherence to physical laws
  • Lower computational costs
For HPC centers facing ever-growing demand for simulation resources, these efficiencies could become increasingly valuable.
 
Rather than replacing supercomputers, physics-informed AI extends its reach by transforming expensive simulation results into reusable computational knowledge.

A new role for supercomputing

The implications extend well beyond photonics.
 
Many of the grand challenges tackled by modern supercomputers involve simulations that are both computationally expensive and physics-rich:
  • Materials discovery
  • Semiconductor design
  • Aerospace engineering
  • Energy systems
  • Climate science
  • Quantum device development
  • Advanced manufacturing
Each field generates vast quantities of simulation data that could potentially be transformed into intelligent surrogate models.
 
In this vision, supercomputers evolve from engines of calculation into engines of knowledge generation.
 
Their role shifts from repeatedly solving the same equations toward training AI systems capable of applying that knowledge across millions of new scenarios.

The curious future of scientific discovery

What makes the Chalmers work particularly fascinating is that it offers a glimpse of a future where AI and supercomputing become inseparable partners.
 
The next scientific breakthrough may not come from a larger neural network alone, nor from a faster supercomputer operating in isolation.
 
Instead, it may emerge from systems in which supercomputers teach AI the fundamental laws governing the physical world, and AI returns the favor by making that knowledge instantly accessible to researchers.
 
The Chalmers “digital super brain” represents an early example of this emerging paradigm, a future where computational science is accelerated not only by more processing power, but by machines that learn directly from physics itself.
 
For the HPC community, that may be the most significant discovery of all.
Featured

From DNA to digital twins: New MDNA framework brings AI supercomputing closer to whole-cell simulation

Tyler O'Neal, Staff Editor LATEST
The next major breakthrough in computational biology may not come from a new supercomputer, but from the software that allows scientists to harness one.
 
Researchers have introduced MDNA, a new open-source molecular modeling framework designed to generate, manipulate, and analyze complex DNA structures with unprecedented flexibility. While the software itself is a biological modeling tool, its broader significance lies in how it could accelerate large-scale molecular simulations, AI-driven biological discovery, and ultimately the long-standing ambition of constructing digital twins of living systems.
 
At a time when exaflops computing is transforming fields ranging from climate science to astrophysics, biology is increasingly becoming one of the most computationally demanding scientific disciplines. DNA is no longer viewed simply as a sequence of genetic letters. It is a dynamic three-dimensional structure whose geometry, interactions, and modifications influence everything from gene expression to disease progression.
 
Understanding those behaviors requires simulation at an extraordinary scale.

Building better starting points for supercomputing

Modern molecular dynamics simulations can model systems containing millions, or even billions, of atoms. Researchers have already demonstrated billion-atom DNA simulations on leadership-class supercomputers, revealing how genes fold, interact, and regulate biological activity.
 
However, one persistent challenge has been constructing biologically realistic DNA configurations suitable for large-scale simulation.
 
MDNA addresses that bottleneck.
 
The framework enables researchers to generate DNA structures with arbitrary shapes using spline-based modeling techniques, while also supporting biologically important modifications such as DNA methylation, Hoogsteen base-pair transitions, and non-canonical nucleotide configurations. By integrating structure generation and structural analysis within a single Python-based workflow, the software streamlines the creation of simulation-ready DNA systems.
 
The result is a platform that reduces the time required to translate a biological hypothesis into a computational experiment.

Bridging AI and molecular simulation

One of the most compelling aspects of MDNA is its compatibility with many computational tools already used across the molecular simulation community.
 
The software integrates with established platforms such as OpenMM, MDAnalysis, MDTraj, oxDNA, Bio3D, HTMD, and PLUMED, making it easier to connect AI-generated molecular designs with high-performance simulation workflows. According to the authors, the goal is not merely to construct DNA structures, but to enable a complete computational ecosystem for studying DNA-protein interactions and molecular dynamics.
 
This arrives at a pivotal moment for computational biology.
 
Artificial intelligence is increasingly being used to design biological molecules, predict molecular structures, and explore vast biochemical design spaces. Recent advances have demonstrated AI-driven approaches to genetic circuit design and biomolecular engineering, generating datasets and candidate structures at scales impossible for human researchers alone.
 
Yet AI predictions are only the beginning.
 
Before a new biological design can be trusted, it often must be validated through detailed molecular simulations that capture physical behavior at atomic resolution. These simulations frequently require the resources of modern supercomputing facilities.
 
MDNA provides a bridge between those two worlds.

Toward digital twins of biology

The implications extend well beyond DNA modeling.
 
Scientists increasingly envision a future in which entire biological systems can be represented as computational “digital twins,” virtual counterparts capable of predicting molecular behavior, disease progression, or therapeutic outcomes before laboratory experiments are performed.
 
Recent projects have mapped the four-dimensional organization of the human genome with unprecedented detail, identifying hundreds of thousands of genomic interactions across time and space.
 
At the same time, researchers are developing computational frameworks capable of simulating cellular processes, cancer evolution, and molecular communication networks.
 
Such ambitions depend on accurate structural models as foundational inputs.
 
MDNA represents one piece of that larger puzzle: a software layer capable of generating realistic DNA architectures that can be incorporated into increasingly sophisticated simulations.

The road to whole-cell simulation

Perhaps the most inspiring aspect of the work is what it suggests about the future.
 
For decades, biologists have dreamed of creating computational models capable of simulating entire living cells. Achieving that goal requires integrating DNA, proteins, RNA, membranes, molecular machinery, and environmental interactions into unified computational frameworks.
 
Exaflops supercomputers are beginning to provide the raw computational horsepower needed for such efforts. Yet hardware alone is insufficient. Researchers also require software capable of building, organizing, and analyzing the immense biological structures that those machines will simulate.
 
MDNA helps fill that gap.
 
By simplifying the construction of highly detailed DNA systems and integrating them with modern simulation ecosystems, the framework contributes to the growing software infrastructure underpinning next-generation computational biology.

A new era for computational life sciences

While the history of supercomputing is often defined by raw hardware power, scientific progress increasingly relies on the sophisticated software frameworks that translate that capacity into actionable insight.
 
MDNA exemplifies this shift: although it may not be the largest or most intensive platform, its value lies in its ability to bridge the gap between AI-driven discovery and large-scale molecular simulation.
 
By simplifying the complexity of DNA modeling, MDNA provides a vital tool for the long-term goal of building biological digital twins.
 
As we enter the exaflops era, such software will be indispensable, proving that while the future of life sciences is written in DNA, it will be mapped through the power of advanced computational modeling.
Microscope image of a semiconductor-integrated spintronic test chip developed by researchers at Tohoku University and NIST. The device demonstrates the first silicon-integrated probabilistic bit (p-bit), a key building block for future large-scale probabilistic computers designed for AI and optimization workloads.
Microscope image of a semiconductor-integrated spintronic test chip developed by researchers at Tohoku University and NIST. The device demonstrates the first silicon-integrated probabilistic bit (p-bit), a key building block for future large-scale probabilistic computers designed for AI and optimization workloads.
Featured

Silicon spintronics brings the P-computer closer to reality

Deck LATEST

Tohoku University and NIST demonstrate the world’s first semiconductor-integrated spintronic p-bit, opening a path toward million-bit probabilistic computers for AI and optimization

Efforts to develop computing architectures that surpass conventional CMOS and modern AI accelerators have just advanced significantly. A team from Tohoku University and the National Institute of Standards and Technology (NIST) has created the world’s first semiconductor-integrated spintronic probabilistic bit (p-bit), fabricated directly on a silicon chip using standard semiconductor manufacturing methods. This breakthrough tackles one of the main challenges in probabilistic computing: scalability.
 
For the high-performance computing community, this result is especially compelling. It marks a shift from laboratory prototypes using discrete components to integrated architectures that could ultimately support large-scale AI and optimization workloads.

A different path beyond Moore’s Law

As AI models continue to grow, traditional processors increasingly struggle with the energy costs associated with searching enormous solution spaces. While quantum computing remains a promising long-term approach, researchers worldwide are also exploring alternative architectures that can handle probabilistic calculations more efficiently.
 
One such architecture is the probabilistic computer, or p-computer.
 
Unlike conventional bits that are fixed at either 0 or 1, p-bits fluctuate stochastically between the two states. Their probability distributions can be controlled and correlated with neighboring p-bits, allowing entire networks to explore many possible solutions simultaneously. This makes p-computers particularly attractive for combinatorial optimization, sampling, machine learning, and inference problems.
 
Previous demonstrations of spintronic p-computers relied on separate spin devices connected to external control electronics through cables. Those systems successfully validated the concept but were limited to roughly 100-bit-scale experiments and offered little opportunity for the kind of integration required for practical computing systems.

Integrating spintronics directly into silicon

The new work changes that equation.
 
The research team fabricated a p-bit circuit directly on a silicon substrate by combining advanced semiconductor manufacturing techniques in the United States with spintronic device fabrication performed at Tohoku University. The resulting device integrates CMOS circuitry with a superparamagnetic tunnel junction whose magnetic state fluctuates naturally due to thermal effects.
 
The prototype was fabricated using a 130-nanometer CMOS process and experimentally verified to exhibit the expected probabilistic input-output behavior required for p-bit operation. The work was reported in IEEE Electron Device Letters https://ieeexplore.ieee.org/document/11535457/.
 
While a single p-bit may appear modest compared with modern processors containing billions of transistors, the accomplishment is significant because it demonstrates a manufacturing pathway compatible with semiconductor-scale integration.
 
According to the researchers, this foundational technology could eventually enable systems containing on the order of one million p-bits, representing a dramatic leap beyond current demonstrations.

Why supercomputing researchers should pay attention

For the HPC community, probabilistic computing occupies an increasingly interesting niche between traditional computing and quantum computing.
 
Many computational science problems involve searching extremely large solution spaces:
  • Protein folding and molecular sampling
  • Logistics and routing optimization
  • Bayesian inference
  • Machine learning training and inference
  • Statistical physics simulations
  • Financial risk modeling
These workloads often consume enormous amounts of compute time on today’s GPU-powered supercomputers.
 
P-computers are not quantum computers and do not rely on fragile quantum coherence. Instead, they exploit naturally occurring randomness in physical devices, operate at room temperature, and use mature semiconductor manufacturing techniques.
 
If large-scale p-bit arrays become practical, they could emerge as specialized accelerators analogous to GPUs or AI tensor processors, targeting classes of problems where stochastic search and probabilistic inference dominate computational cost.

The spintronics connection

The work also highlights the growing importance of spintronics as a post-CMOS technology platform.
 
Spintronic devices use both the charge and the spin of electrons, enabling information processing mechanisms that differ fundamentally from traditional transistor logic. Technologies such as MRAM have already demonstrated that spintronic devices can be manufactured within semiconductor production flows.
 
The ability to combine CMOS circuitry and stochastic magnetic tunnel junctions on the same silicon substrate suggests that future p-computers could leverage existing semiconductor infrastructure rather than requiring entirely new fabrication ecosystems.
 
That compatibility may prove decisive in determining whether probabilistic computing remains a research curiosity or evolves into a deployable computing technology.

A new computing landscape

The history of computing is filled with architectures that appeared promising but never escaped the laboratory. What makes this announcement noteworthy is not merely the demonstration of another novel device, but the demonstration of a manufacturable pathway.
 
The researchers have effectively shown that spintronic p-bits can be integrated into semiconductor processes rather than attached as external experimental components. That shift transforms probabilistic computing from a proof-of-concept architecture into a technology with a plausible scaling roadmap.
 
For supercomputing researchers watching the search for post-Moore computing platforms, the question is no longer whether spintronic p-bits can operate on silicon. That has now been demonstrated.
 
The more interesting question is what happens when hundreds of thousands, or even millions, of them begin working together.