Neural networks have transformed how machines find patterns in data, from recognizing faces in photos to predicting the shapes of proteins. So far, all of this progress has been made on ordinary classical computers, but with quantum computers now edging into practical use, there is a real possibility that neural networks could tap into distinctly quantum effects and operate in ways that classical machines never could. So far, however, neural networks have proven far more difficult to run on quantum hardware.

Through new research published in Physical Review Letters, Djamil Lakhdar-Hamina and colleagues at the University of Maryland, College Park, have built a neural network that runs on two different types of quantum computer, allowing them to test directly whether these systems can live up to their theoretical promise.

Elusive quantum advantage
A neural network is built from layers of simple units, each taking in signals and passing on an output depending on what it receives. To train a network, the connections between these units are adjusted until the network reliably produces the right answer for a given task.

In the quantum world, a similar structure can be built using qubits: the basic unit of quantum information, whose measurement outcomes stand in for the signals passed between layers. Researchers have long suspected that quantum versions of these networks could offer genuine advantages over classical ones, perhaps by exploiting quantum uncertainty. However, very few of these ideas have actually been tested on physical devices.

Researchers at Rensselaer Polytechnic Institute (RPI) have found that today's leading artificial intelligence tools for predicting protein structures routinely generate results that are physically and chemically impossible, exposing critical blind spots in how AI is being applied across scientific research. The work, published in the Proceedings of the National Academy of Sciences, serves as a cautionary reminder that AI still requires human oversight and physics-based verification to produce reliable results in the lab.

The paper, authored by George I. Makhatadze, professor of biological sciences and Constellation Endowed Chair at RPI, evaluated widely used deep learning tools for predicting how flat sequences of amino acids fold into the three-dimensional structures that determine a protein's function. Makhatadze found that these tools frequently overlook the underlying scientific rules of protein folding—and, notably, that every tool tested rated its own accuracy higher than the results warranted.

"The major conclusion of the paper essentially is: trust but verify," Makhatadze explained. "You have to verify [AI outputs] using physics-based methods."

Where the models break down
AI has become indispensable for analyzing the massive data sets used to predict protein folds. For example, Google's DeepMind AI laboratory—known for AlphaFold2—shared the 2024 Nobel Prize in Chemistry for its contributions to protein structure prediction.

But according to Makhatadze's work, AlphaFold2 and RoseTTAFold2—a similar deep learning-based prediction platform developed at the University of Washington—both produced "implausible structures for variant sequences" by "[prioritizing] statistical patterns over the underlying thermodynamic principles of folding." Both tools are trained on evolutionary data and structural databases.

"AlphaFold is considered the gospel of the field," Makhatadze said. "It is very good, and it does many things well. But occasionally it makes mistakes, because there simply isn't enough of the right kind of data in the model yet."

Makhatadze found fewer scientific impossibilities in a different class of tools—"transformer-based protein language models" that rely on protein sequences rather than structural data. Tools in this class included OmegaFold and the Meta-developed ESMFold. However, neither category of model performed well when proteins contained ionizable residues, meaning amino acid side chains that can gain or lose a proton depending on their surrounding environment.

Researchers have uncovered a surprising mechanism that allows plants to carefully coordinate the formation and growth of new leaf parts. The study shows that the hormone auxin regulates different phases in organ formation by oppositely affecting the activity of another hormone, gibberellin, to trigger the formation of new leaf structures before reversing course and boosting gibberellin to drive their expansion. The findings offer new insight into how plants build complex organs and could eventually help scientists develop crops with improved growth and architecture.

A plant's ability to produce leaves, flowers and other organs depends on precise location and timing. It must first determine where a new structure will form before allowing it to expand. Now, researchers have uncovered the molecular switch that coordinates these two steps, revealing how plants carefully alternate between putting on the brakes and stepping on the accelerator during organ development.

If you own a strong enough telescope, you might be able to witness history on Wednesday: A dead rocket is going to crash into the moon. On August 5, at around 6:34 am UTC, a spent SpaceX Falcon 9 upper stage is expected to hit the moon's sunlit western limb near Einstein crater at more than 5,400 mph. If it unfolds as predicted, the crash could throw up a plume of debris bright enough to briefly see from Earth with the right equipment.

That would be a first. No impact flash has ever been recorded on the sunlit face of the moon, and that's exactly where the Falcon 9 crash is forecast to happen, kicking up plumes of dust that could stand out against the blackness of space. The findings are based in part on two new preprint studies.

That includes one posted July 27 to arXiv and led by William Jo, a doctoral candidate at the University of Texas at Austin's Cockrell School of Engineering, that forecasts just how big the impact could be. To predict the plume, Jo ran the crash through a high-resolution physics simulation that allowed him to model what would happen when 3,900 kilograms of hollow metal hit the lunar surface. That’s different from solid meteorites making impact.

"It's like an empty eggshell, because it had all the fuel in it, and there is a rocket engine at one end that's denser," David Goldstein, an aerospace engineering professor at UT Austin who supervised the work, says of the Falcon 9.

Rather than burrowing in like a cannonball, the shell will collapse from its edges inward, throwing up a broad, low curtain of soil spreading as far as 183 kilometers wide as well as a thin, faster spike nearly straight up. All told, the crash is forecast to displace about 12,700 kilograms of debris.

Graphite, the carbon core of a humble No. 2 pencil, is also an essential component in technologies such as batteries, smartphones, laptops and industrial power equipment. Today, nearly all of this critical mineral must be mined and processed and, in the United States, imported.

But now, researchers at the U.S. Department of Energy's Lawrence Berkeley National Laboratory (Berkeley Lab), UC Berkeley and Estonia's National Institute of Chemical Physics and Biophysics have shown a promising way to convert waste carbon pulled from the air into graphite, opening up a potential alternative to mining. The work was published recently in the journal Nature Communications.

Researchers built a custom microscope setup to watch a process known as molten-salt electrolysis, which uses electricity and hot liquid salts to turn carbon dioxide into solid carbon. For the first time, researchers were able to watch the process in real time inside corrosive molten salts heated to 500°C (932°F) while the system was running.

The observations answered a decades-old question about how the reaction occurs at the molecular level, revealing an unexpected two-step process. The researchers also found that the basic chemical reaction remained the same even when they changed the materials used for the electrodes and molten salts. Because different materials produce different carbon structures, scientists should be able to tune the process to make valuable carbon products, with the goal of making battery-grade graphite.

On Aug. 12, a total solar eclipse will cross Greenland, Iceland, and Spain — and NASA science will be there! ☀️🌑🔭

We're flying high-altitude jets and launching scientific balloons to study the Sun and the eclipse's effects on us: go.nasa.gov/4x5qvqP

Chinese researchers have revealed a new technology that could soon enable drones to charge mid-flight using high-powered lasers.

The scientists built a prototype of the system using a model of a drone and attached a receiver that works similarly to a solar cell. Fixed to the underside of the wing, the receiver successfully converted the energy from the laser beam into electricity to power the aircraft’s propellers.

"Previous studies largely focused on the materials or the device itself," study senior author Jianhua Han, a researcher at the Civil Aviation University of China, said in a statement. "We wanted to think beyond the laboratory, to how the system could actually be integrated into an aircraft, cooled during operation, and made compatible with flight. It isn’t just a materials science problem; it’s an engineering one."

New blood tests for Alzheimer's disease aim to catch the condition in its earliest stages. Scientists hope that, someday, by spotting very early signs of dementia and enabling earlier treatments, these tests could significantly improve patients' prognoses.

But how well do current blood tests work for that purpose?

At the Alzheimer's Association International Conference in London on July 12-15, Live Science spoke with researchers and clinicians who study and use blood tests for Alzheimer's disease. Here's their take on the current state of these tests and their practicality for real-world use

A new visualization of Earth looks nothing like the Blue Marble seen from space.

NASA has released a new geoid model showing the shape of Earth’s gravitational field. The slight, irregular bulges and depressions are caused by variations in the distribution of mass within the planet. It can also be imagined as the surface a global ocean would take if it were determined solely by Earth’s own gravity and rotation.

The visualization is an exaggerated geoid, with heights multiplied by a factor of 10,000 in order to show the variations in the gravitational field. NASA created this lumpy version of Earth using models and data from satellites, which can detect minute differences in gravity at different locations. The dataset stretches back 15 years and includes a billion observations. NASA also shared an interactive 3D model of the geoid that you can explore here, as well as a longer video.