Unsolved Physics and Chemistry Problems Today, and What AI Has Done

The biggest open problems in physics are quantum gravity, the nature of dark matter and dark energy, why matter beat antimatter, and what a quantum measurement really is. In chemistry they include catalysis, nitrogen fixation, better batteries and materials discovery. AI has delivered one Nobel level result, protein structure prediction, and many claims that needed correction.

This post walks through each problem using primary sources: what is known, what would count as solved, and where AI has helped or overreached. It also covers how scientists check a claimed discovery, because the same habits (evidence, provenance, independent replication) are what make software trustworthy. One thing up front: Safeguard's models do not solve physics or chemistry problems. They are single purpose security models, and Safeguard's own benchmarks page says a general frontier model will out-write them on general tasks.

Key takeaways

• Physics has a theory that works almost everywhere and still cannot say what most of the universe is made of. According to NASA, ordinary matter is about 5 percent of the universe, dark matter about 27 percent and dark energy about 68 percent.
• The matter and antimatter imbalance is still unexplained. CERN describes it as roughly one extra particle of matter per billion antiparticles after the Big Bang. LHCb reported the first observation of CP violation in baryon decays in Nature on 16 July 2025.
• Fusion ignition is real and repeatable at Lawrence Livermore's National Ignition Facility: 3.15 MJ out from 2.05 MJ of laser energy on 5 December 2022, and 8.6 MJ on 7 April 2025. That is target gain, not a power plant.
• Room temperature superconductivity remains unconfirmed. LK-99 (2023) turned out not to be a superconductor, and Nature retracted a near ambient superconductivity paper on 7 November 2023.
• AI's clearest win is protein structure prediction. The 2024 Nobel Prize in Chemistry went half to David Baker for computational protein design and half to Demis Hassabis and John Jumper for AlphaFold.
• AI materials discovery claims have needed public correction. An autonomous lab paper published in Nature in 2023 received an author correction on 19 January 2026 clarifying that "new" meant new to the prediction platform, not new to science.
• Verification is the common thread: a result counts when someone else can check the inputs and rerun the process. Software works the same way, through provenance, signatures and reproducible builds.

What does "unsolved" mean in physics and chemistry?

"Unsolved" covers three different kinds of gap, and mixing them up causes most of the confusion in popular coverage.

The first is an explanation gap: we can measure something precisely but have no accepted theory for why it is so. Dark energy is the cleanest example. NASA's summary of what dark energy is starts with a short answer: we do not know. We can measure its effect on cosmic expansion, but not its nature.

The second is a theory gap: two theories that each work extremely well do not fit together. General relativity and quantum mechanics are the textbook case.

The third is an engineering or synthesis gap: the physics is understood well enough, but nobody has made the thing work at the needed scale, cost or conditions. Fusion energy, ambient condition ammonia synthesis and solid state batteries sit here.

Each gap closes differently: a confirmed new prediction, a unified framework that makes testable predictions, or a device that works repeatedly under independent measurement.

Particle physics adds a statistical bar. In its July 2025 Nature paper, the LHCb collaboration describes an earlier 3.1 standard deviation effect as a hint "requiring further confirmation" and its new 5.2 standard deviation result as an observation. That five sigma convention is why many exciting results stay in the "hint" column for years.

Open problems at a glance

Physics: the big open questions

Why is quantum gravity still unsolved?

General relativity, Einstein's 1915 theory of gravity, describes gravity as the curvature of spacetime. Quantum theory describes the other forces and matter at small scales. Each has passed every test thrown at it in its own domain. The trouble is that they use incompatible pictures. General relativity treats spacetime as smooth and classical; quantum theory treats everything as probabilistic and discrete. Where both matter at once, inside black holes or at the first instant of the Big Bang, the combination breaks.

The ALPHA collaboration put it plainly in its 2023 Nature paper on antimatter and gravity: singularities in general relativity and the lack of a quantum theory of gravity suggest our picture is incomplete. CERN's own explainer frames a related puzzle, why gravity is so much weaker than the other forces, and notes that a small fridge magnet can out-pull the whole Earth. One possibility CERN describes is that part of gravity spreads into extra spatial dimensions, which could also bear on why the universe is expanding faster than expected. No experiment has confirmed extra dimensions.

The core difficulty is data. Candidate theories can be mathematically rich and still make no prediction anyone can test with current experiments. AI tools can help check algebra or search spaces of equations, but a theory still has to predict something an experiment confirms.

What are dark matter and dark energy, and what changed with DESI?

Dark matter is the name for mass we infer from gravity but cannot see. NASA describes it as making up most of the mass in galaxies and galaxy clusters, about 27 percent of the universe, and says scientists have been trying to identify it for almost 100 years. NASA adds that it is thought to shape the cosmos, organising galaxies on large scales. CERN's dark matter page makes the key limitation plain: we can only detect it through its gravitational effects. No candidate particle has been confirmed.

Dark energy is the name for whatever is making the expansion of the universe speed up. According to NASA, the acceleration started about nine billion years after the Big Bang, dark energy was discovered in the late 1990s, and it makes up about 68 to 70 percent of the universe. The standard model of cosmology treats it as a constant, the cosmological constant, written Lambda in "Lambda CDM".

The interesting news is from the Dark Energy Spectroscopic Instrument (DESI). On 19 March 2025 the collaboration released results from its second data release, based on three years of observations and more than 14 million galaxies and quasars. The headline paper (arXiv 2503.14738) reports that a dark energy whose equation of state changes over time fits the data better than a constant. Combining DESI with cosmic microwave background data, that preference is 3.1 standard deviations. Adding supernova samples moves it to between 2.8 and 4.2, depending on which supernova sample is used.

That is a serious hint, below the five sigma bar, and it shifts with the external dataset combined. If it holds, dark energy is not a simple constant. If it fades, that is science working too.

Why is there more matter than antimatter?

Paul Dirac's 1928 equation predicted antimatter: for every particle, a partner with opposite charge. Particles and antiparticles are made in pairs and annihilate on contact. CERN's explainer lays out the problem: if the early universe made them in equal amounts, today's universe should hold equal amounts of both, or nothing but leftover radiation. Instead it holds matter and almost no antimatter. CERN's current best description is that some unknown mechanism created roughly one extra particle of matter for every billion antiparticles.

One required ingredient is CP violation, a difference in how matter and antimatter behave. It has been seen in mesons (particles made of a quark and an antiquark). On 16 July 2025, LHCb published in Nature the first observation of CP violation in baryons, the three quark family that includes protons and neutrons. Studying decays of the beauty baryon called Lambda b into a proton, a kaon and two pions, they measured an asymmetry of about 2.45 percent at 5.2 standard deviations.

That result fills in a gap in the Standard Model's checklist. It does not solve the asymmetry. The paper itself frames the observation as confirming that baryonic matter and antimatter behave differently, which the Standard Model predicts. The open question is still where the much larger imbalance in the universe came from.

ALPHA at CERN tested whether antimatter falls up. Its Nature paper of 27 September 2023 reports that antihydrogen released from a magnetic trap behaved consistently with ordinary gravitational attraction, ruling out repulsive "antigravity" in that setup.

What is the measurement problem in quantum mechanics?

Quantum theory predicts measurement probabilities with extraordinary accuracy but does not settle what a measurement is. Between measurements a system evolves as a superposition; at measurement we see one outcome. Is that a physical collapse, a branching of worlds, hidden information, or an update of our knowledge?

The Stanford Encyclopedia of Philosophy's entry on philosophical issues in quantum theory opens by saying there is no consensus among physicists or philosophers on what the success of quantum theory tells us about the world. The approaches it surveys, hidden variable theories, collapse theories and others, add different things to a shared operational core.

Experiments have narrowed the field. The 2022 Nobel Prize in Physics, announced on 4 October 2022, went to Alain Aspect, John Clauser and Anton Zeilinger for experiments with entangled photons that established violations of Bell inequalities. Those experiments rule out a broad class of local hidden variable explanations. They do not pick one interpretation. The interpretations agree on the predictions of the operational core, which is why experiment alone has not settled the question.

Is room temperature superconductivity real?

A superconductor carries current with zero resistance and expels magnetic fields (the Meissner effect). Every known superconductor needs either very low temperatures or extreme pressures. One that worked at room temperature and normal pressure would change power grids, magnets and electronics. That is why claims of one spread fast, and why they get checked hard.

Two recent episodes show how the checking works.

LK-99. In July 2023 a preprint on arXiv (2307.12008) claimed the first room temperature, ambient pressure superconductor, with a critical temperature at or above 400 K (127 degrees Celsius), in a modified lead apatite material. Labs around the world tried to replicate it within weeks. By 16 August 2023, Nature's news team reported that LK-99 is not a superconductor: replication work had pieced together why the material showed superconducting-like behaviour and clarified its actual properties.

The lutetium hydride paper. Nature published a paper claiming near ambient superconductivity in nitrogen doped lutetium hydride. On 7 November 2023 Nature retracted it. The retraction note says eight of the authors asked for it, because the paper did not accurately reflect "the provenance of the investigated materials", the measurements taken or the data processing applied. Separately, Nature's own investigation and post publication review found concerns about the electrical resistance data to be credible, substantial and unresolved.

Notice the word provenance in a chemistry retraction. Where a sample came from, how it was made and how the raw data became the published figure are exactly the questions that decide whether a result stands. Software has the same questions about where a binary came from.

So no confirmed room temperature, ambient pressure superconductor exists. Any claim needs independent replication of both zero resistance and the Meissner effect.

Did fusion ignition solve fusion energy?

No, but it crossed a real scientific threshold. On 5 December 2022, the National Ignition Facility at Lawrence Livermore National Laboratory delivered 2.05 MJ of laser energy to a fuel target and got 3.15 MJ of fusion energy out. The US Department of Energy and NNSA announced it on 13 December 2022 as the first controlled fusion experiment to produce more energy from fusion than the laser energy used to drive it. DOE's own release added that many advanced science and technology developments are still needed for practical inertial fusion energy.

LLNL has repeated ignition since. Its "Achieving Fusion Ignition" page lists these milestones:

LLNL defines target gain as energy yield divided by the energy on target. That ratio leaves out the electricity the laser system draws to produce those megajoules, and it says nothing about converting fusion energy into electricity, breeding fuel, or firing many times a second. Those are the engineering gaps between ignition and a power plant.

AI has a role on the magnetic confinement side. In a Nature paper published on 16 February 2022, Degrave and colleagues described a deep reinforcement learning controller that learned to command the full set of magnetic control coils of the TCV tokamak (Tokamak a Configuration Variable) and held a range of plasma shapes, including two separate plasma "droplets" at once. That is a control result, not net energy.

Chemistry: the big open questions

Did AlphaFold solve the protein folding problem?

For single protein chains, mostly yes, and the Nobel committee said so. On 9 October 2024 the Royal Swedish Academy of Sciences awarded the Nobel Prize in Chemistry half to David Baker "for computational protein design" and half jointly to Demis Hassabis and John Jumper of Google DeepMind "for protein structure prediction". The press release describes structure prediction from amino acid sequence as a 50 year old problem, and says AlphaFold2, presented in 2020, has been used to predict the structure of virtually all of the roughly 200 million proteins researchers have identified, and has been used by more than two million people from 190 countries.

The Nobel committee's popular science background tells the verification story too. Since 1994 the Critical Assessment of Protein Structure Prediction (CASP) has run as a blind test: entrants get sequences whose structures have just been solved experimentally but not released, and predictions are scored against the hidden answers. Before 2018 the best predictions reached about 40 percent accuracy on the CASP scale. AlphaFold reached almost 60 percent in 2018, still short of the roughly 90 percent needed for success. AlphaFold2, presented in 2020, is the version the Nobel committee credits with the breakthrough.

That blind test is the important part: the AI was judged against held back experimental data, not its developers' say so.

What remains open? AlphaFold 3, described in Nature on 8 May 2024, extends prediction to complexes of proteins with nucleic acids, small molecules, ions and modified residues, with large accuracy gains on protein and ligand interactions. Even so, a structure prediction is a static snapshot. How proteins move, how they misfold in disease, and how folding happens step by step in the cell are still active research. Predicting a structure is also not the same as knowing what a drug will do in a body.

Can we design new proteins from scratch?

Yes, and that is the other half of the 2024 prize. According to the Nobel popular information, Baker's group designed Top7, a 93 amino acid protein with a fold not seen in nature, and published it in 2003. They drew the target structure first and used their Rosetta software to find a sequence that would fold into it. Since then the lab has designed proteins that bind targets, assemble into cages and act as sensors.

Design flips the prediction problem: which sequence gives the shape you want? The open problems are in function, such as enzymes as efficient as natural ones. Every design still has to be made in a lab and measured.

Why is nitrogen fixation still a hard catalysis problem?

Plants need nitrogen in chemical compounds, while the air is mostly unreactive nitrogen gas. Fritz Haber developed a method for making ammonia from nitrogen and hydrogen around 1913, using a catalyst under controlled temperature and pressure, and received the 1918 Nobel Prize in Chemistry (presented in 1919). The industrial version, the Haber Bosch process, dominates industrial ammonia production today, as the paper below notes.

The open problem is doing it with less energy. A 2019 Nature paper by Andersen and colleagues describes Haber Bosch as energy intensive and electrochemical synthesis under mild conditions, powered by renewable electricity, as an attractive alternative. Then it delivers the uncomfortable part: most reported electrochemical results produced so little ammonia that it could not be firmly attributed to nitrogen fixation rather than contamination. Ammonia can come from air, from human breath, from the membranes in the cell, or from nitrogen compounds in the gas supply or even the catalyst itself.

Their fix was a protocol: feed isotopically labelled nitrogen 15 and quantify it, so the product can only have come from the supplied gas, and run controls that eliminate each contamination source. That is provenance and controls applied to a molecule.

AI screening can narrow catalyst candidates. Measurement under a rigorous protocol still decides.

What is still unsolved in battery chemistry?

The 2019 Nobel Prize in Chemistry went to John Goodenough, Stanley Whittingham and Akira Yoshino for developing the lithium ion battery, whose foundations, according to the Nobel press release, were laid during the 1970s oil crisis. Lithium ion cells now power phones, laptops and electric cars and store energy from solar and wind.

What remains open is a set of trade offs: more energy per kilogram, faster charging, longer life, less scarce material and better safety. Solid electrolytes are one long running hope, and the 2023 GNoME paper, discussed below, lists solid electrolyte candidates among its outputs. A predicted stable crystal still has to survive synthesis, conductivity measurement and real cycling before it becomes a cell.

Can AI discover new materials, and how are the claims checked?

This is the most instructive story in the post, because it shows a strong AI result, an overclaim and a correction in public view.

The prediction side. On 29 November 2023, Google DeepMind published "Scaling deep learning for materials discovery" in Nature. The abstract says graph networks trained at scale found 2.2 million structures below the current convex hull (that is, predicted to be stable against decomposition into known phases), and that 736 of them had already been independently realised by experiment. DeepMind's announcement describes 380,000 of them as stable materials. The paper also produced learned interatomic potentials used for molecular dynamics.

The synthesis side. The same day, Nature published a paper by Szymanski and colleagues introducing the A-Lab, an autonomous laboratory that combines computation, literature data, machine learning and robotics to synthesise inorganic powders. Its current abstract says that over 17 days of continuous operation it realised 36 compounds from a set of 57 targets.

The challenge. On 7 March 2024, a team from Princeton and University College London (Leeman, Schoop, Palgrave and colleagues) published a critique in PRX Energy. They examined all 43 synthetic products that they say the A-Lab paper reported as new, identified four common shortfalls in the analysis, and concluded that no new materials had been discovered in that work. Two of their points are general: automated Rietveld analysis of powder X-ray diffraction data is not yet reliable, and predictions often ignore disorder, where elements share crystal sites and the "new" compound is really a known alloy or solid solution.

The correction. On 19 January 2026, Nature published an author correction to the A-Lab paper. The authors say concerns were raised about unambiguous structure identification and about novelty claims. They state that the novelty claims were meant to say the materials were new to the prediction platform, not necessarily new to science, and updated the text. They manually re-analysed the diffraction patterns, confirmed 36 of 40 reported successes, judged four inconclusive, and removed one compound that had mistakenly been included in the training data. The re-analysis was itself peer reviewed after publication.

Nobody had to take anyone's word. The data were available, other chemists re-analysed them, and the record was corrected. Training data contamination and "new to our system versus new to the world" are failure modes anyone evaluating AI output should recognise, including security teams reviewing AI found vulnerabilities.

What role does AI actually play in physics and chemistry?

Across the sources above, AI contributes in five distinct ways. They deserve different levels of trust.

What AI has not done, according to any primary source we found, is produce a new fundamental theory, such as a theory of quantum gravity or an explanation of dark energy, that experiments then confirmed. Researchers also use general language models for literature search, analysis code and checking algebra. This post makes no claim about any specific model's scientific ability and no ranking of models; whatever a model writes is a claim that needs the checks below.

Where does quantum simulation fit?

Simulating molecules and materials exactly is hard for classical computers because the quantum state of many interacting electrons grows exponentially with system size. Quantum computers are proposed partly to address that. In its 9 December 2024 Nature paper on the Willow processor, Google Quantum AI lists quantum chemistry and quantum simulation among the speedups quantum computing promises, then notes that quantum information is fragile and error prone.

The result itself is about error correction: a 101 qubit, distance 7 surface code reached 0.143 percent error per cycle, with logical errors suppressed by a factor of about 2.14 each time the code distance grew by two. That is a step toward useful quantum simulation, not a chemistry result. Check any claim that a quantum computer "solved" a chemistry problem against the paper and the best classical method.

Can Safeguard's models solve physics or chemistry problems?

No. Safeguard's Griffin, Eagle and Lion models are trained for software security work: reachability analysis, patch synthesis, triage and compliance evidence mapping. Safeguard's public benchmarks page says plainly that a frontier general purpose language model will out-write any Safeguard model on poetry, summarisation, math word problems and general code completion, and that the lineup is single purpose by design.

The overlap with this topic is method, not subject matter. A reachability result should be a checkable claim: here is the call path from your code to the vulnerable function, with the evidence. A patch should come with a test run. That is the same discipline as a CASP blind test or an isotope labelled control, applied to code.

Step by step: check a discovery claim, then check your software the same way

Here is a practical checklist for reading a science headline, followed by the software equivalent with real commands.

1. Find the primary source and its current version

Locate the paper, preprint or lab release, not a summary. Check arXiv version history, and look for "Author Correction", "Retraction Note" or "Change history" on the journal page, as the A-Lab and lutetium hydride papers show.

2. Separate the claim from the framing

Write the claim in one sentence with its number and its uncertainty. "3.1 standard deviations preference for evolving dark energy in one data combination" is a claim. "Dark energy is changing" is framing. "36 compounds from 57 targets, new to the platform" is a claim. "AI discovers new materials" is framing.

3. Ask what independent check exists

Was it a blind test like CASP, replicated by another lab like LK-99, or re-analysed from raw data like A-Lab? A single team's internal result is a lead, not a conclusion.

4. Look for provenance and controls

Where did the sample, dataset or training set come from? What controls ruled out contamination? The ammonia protocol exists because contamination explained many earlier results. The A-Lab correction removed a compound that had leaked into training data.

5. Apply the same steps to the software you run

Software has primary sources (the source repository and tag), claims (this binary was built from that source), framing (a vendor's release note), and independent checks (signatures, provenance attestations, reproducible rebuilds). The commands below are read only checks you can run today.

Check that a downloaded file matches the published digest:

# Compare against the checksum the project publishes on its release page
sha256sum ./tool-v1.4.2-linux-amd64.tar.gz

Verify npm registry signatures and provenance attestations for an installed project:

cd my-app
npm ci
npm audit signatures

Verify a GitHub artifact attestation for a release binary:

gh attestation verify ./tool-v1.4.2-linux-amd64.tar.gz --owner example-org

Verify SLSA provenance with the reference verifier:

slsa-verifier verify-artifact ./tool-v1.4.2-linux-amd64.tar.gz \
  --provenance-path ./tool-v1.4.2.intoto.jsonl \
  --source-uri github.com/example-org/tool \
  --source-tag v1.4.2

Rebuild from the tagged source and compare bit for bit. This is replication for software:

git clone --branch v1.4.2 https://github.com/example-org/tool.git
cd tool
# Use the project's documented, pinned build environment
make release
sha256sum dist/tool-v1.4.2-linux-amd64.tar.gz
# If digests differ, see exactly where
diffoscope ../tool-v1.4.2-linux-amd64.tar.gz dist/tool-v1.4.2-linux-amd64.tar.gz

The Reproducible Builds project defines a build as reproducible when, given the same source code, build environment and build instructions, any party can recreate bit by bit identical copies of the specified artifacts. That is the software version of independent replication.

6. Inventory what you depend on

You cannot check provenance for components you do not know you have. Generate an inventory and scan it:

# Scan a repository, a container image, or an existing SBOM
safeguard scan --dir .
safeguard scan --image registry.example.com/app:1.4.2
safeguard scan --sbom ./sbom.cdx.json
# Results are written as JSON under ./output

SLSA's build track describes how strong a provenance claim is. In the SLSA 1.1 levels, Build L1 means provenance exists showing how the package was built, Build L2 means that provenance is signed and produced by a hosted build platform, and Build L3 means a hardened build platform that resists tampering during the build. Think of the levels as the difference between a lab notebook, a notarised lab notebook and a notebook kept in a locked, audited room.

Comparing scientific and software verification

Where Safeguard fits, honestly

Safeguard has nothing to do with dark energy, catalysts or plasmas. Its fit with this topic is the verification habit for software. Specifically, Safeguard generates and analyses SBOMs in SPDX and CycloneDX, scans repositories, container images and SBOM files for known vulnerabilities, shows which projects contain a given component and version, and runs reachability analysis so the findings your code actually calls come first. It supports attestation and signing with SLSA, Sigstore and in-toto, and policies and gates evaluated in the platform. When a fixed version exists, autonomous remediation can open a tested fix pull request; it ships switched off. The Safeguard CLI's assistant can edit code and run commands only with your approval.

What it does not do: verify scientific results, run simulations, or act as a general reasoning engine for science. For a research group, the direct use is the ordinary one: knowing which open source libraries your simulation code and lab automation depend on, and whether those builds can be traced and reproduced.

For more on the signing side, see our explainer on what Sigstore is, and for using an inventory under pressure, SBOM incident response.

Common mistakes

• Treating a hint as a discovery. The DESI evolving dark energy preference is 2.8 to 4.2 standard deviations depending on data combination. That is significant and worth following, not settled.
• Reading "ignition" as "fusion power". NIF's target gain compares fusion yield with laser energy on target. It excludes the electricity used by the lasers and everything a power plant needs.
• Believing a preprint because it went viral. LK-99 spread worldwide in days; replication took weeks and said no.
• Assuming peer review is the final check. The lutetium hydride paper passed review at Nature and was retracted months later. Review is a filter, not a guarantee.
• Saying AlphaFold solved biology. It solved, to a Nobel standard, the prediction of static structures from sequence. Dynamics, misfolding and function remain open.
• Taking "new" at face value in AI discovery claims. The A-Lab correction shows "new to the model" and "new to science" are different statements.
• Assuming a quantum computer solved a chemistry problem. Willow's 2024 result is about error correction, not chemistry.
• Skipping provenance because the result looks right. In both labs and build pipelines, a correct looking output from an unknown process is the risky case.

FAQ

What are the biggest unsolved problems in physics today?

Quantum gravity, the identity of dark matter, the nature of dark energy, the matter and antimatter imbalance, and the quantum measurement problem. On the applied side: room temperature superconductivity and practical fusion energy.

Has AI solved any unsolved science problem?

AI's clearest success is protein structure prediction. AlphaFold2 matched experimental accuracy in the blind CASP assessment in 2020, and Demis Hassabis and John Jumper shared half of the 2024 Nobel Prize in Chemistry for it. AI has not produced a confirmed new fundamental theory in physics.

Did the National Ignition Facility achieve net energy?

It achieved target gain above one, starting with 3.15 MJ from 2.05 MJ of laser energy on 5 December 2022. Target gain excludes the electricity that powers the lasers, so it is not net electricity.

Is dark energy changing over time?

Possibly. DESI's March 2025 results prefer time varying dark energy at 2.8 to 4.2 standard deviations depending on the data combined, below the five sigma bar for a discovery.

Why is there more matter than antimatter in the universe?

Nobody knows yet. CERN describes an unknown mechanism that left about one extra matter particle per billion antiparticles. CP violation is a necessary ingredient, and LHCb observed it in baryons for the first time in 2025, but that observation does not explain where the excess came from.

Was LK-99 a room temperature superconductor?

No. Replication efforts in 2023 found it was not a superconductor, as Nature's news team reported on 16 August 2023. No confirmed ambient pressure room temperature superconductor exists.

Did AI discover millions of new materials?

GNoME predicted 2.2 million stable crystal structures, 736 of them already made independently, per its 2023 Nature paper. Predicted is not confirmed, and a related autonomous lab paper was corrected in January 2026.

Can large language models do physics research?

Researchers use them to search literature, write analysis code and check algebra. A model's output is a claim like any other and needs the same checks: primary data, independent replication and controls.

Can Safeguard's AI models solve physics or chemistry problems?

No. Safeguard's models are single purpose security models for reachability analysis, patch synthesis, triage and compliance. Safeguard's own benchmarks page says a general frontier model will out-write them on general tasks such as math word problems.

What do science verification and software supply chain security have in common?

Provenance, independent checks and reproducibility. Labs trace samples and replicate results; software teams list components in SBOMs, sign build provenance and rebuild from source to confirm a binary matches.

Fix it without waiting for upstream

Nothing in this section touches physics or chemistry. It is here because research groups and companies alike run simulation codes, data pipelines and lab automation built on open source libraries, and those libraries ship vulnerabilities and fixes on their own schedule like any other software.

Whatever the vulnerability, most of the risk sits in two gaps. The first is the time between a fix being published and that fix actually running in your production builds. The second, and worse, is the time before any fix exists at all, while a maintainer works out a patch and a release schedule. Waiting through both gaps is a habit, not a requirement. Safeguard gives teams three ways to close them.

1. Autonomous remediation: pull requests, not tickets

When a fixed version exists, Safeguard's Griffin models open a tested, policy gated pull request that moves the affected dependency to the safe version. You can run it on one repository or in bulk across every project that carries the vulnerable version. Remediation covers eight categories, including dependencies, licenses, security posture and findings. Reachability analysis ranks the vulnerable code paths your application actually calls first, so the first pull requests you merge are the ones that matter most. Autonomous remediation ships switched off, with every category set to Manual, so your team decides how much runs on its own and what still needs a human approval.

2. Gold Open Source: the safe version in seconds, free

gold.safeguard.sh is free and needs no login. Search any CVE or package to see affected and fixed versions, CVSS, EPSS, CISA KEV status and a recommended safe version, or use the "Am I affected?" checker. The same data is available through a free read API, README status badges and RSS threat feeds. The safeguard-sh/gold-scan-action GitHub Action can fail a build on a known exploited vulnerability, on any known vulnerability, or never, so a version you have just removed cannot quietly come back in next week's merge.

3. Gold Registry: start from components that are already fixed

registry.safeguard.sh provides hardened, drop in replacements for upstream packages and container images. Every Gold artifact is held continuously to zero exploitable critical or high CVEs, zero malware and a completed license check, with SLSA L3 or higher provenance, a Sigstore signature and a public evidence page you can hand to an auditor. An artifact loses its Gold badge if it regresses. Teams can point their own package directory at the registry so hardened equivalents resolve without changing how developers install anything.

When upstream has no fix yet, or will not ship one, Safeguard first works to get the fix accepted upstream. If the issue is severe or upstream will not move, Safeguard builds a fork, fixes the vulnerability there and publishes a hardened Gold build, so you are not stuck on a maintainer's release schedule. You can also request a hardened, compatibility tested build of a specific package. This fork and patch flow is in early access today, so ask which packages it covers before you plan around it.

One honest limit: when the vulnerable code lives in a vendor appliance, a hosted service or an operating system runtime you do not build yourself, the fix still has to come from that vendor or distribution. The steps above apply to the open source dependencies and container images you build and ship, which is where most teams carry the largest and least tracked share of their exposure.

See how Safeguard would help

None of these demos does science. Each shows a verification habit from this post applied to software, with a human approving each step.

The first demo is a real Safeguard CLI session: asked to fix a lodash CVE, it reads the finding and the advisory, asks before each command, edits package.json to lodash 4.17.21, runs npm install and npm test with approval, prints a summary and pushes nothing. It ends with /permissions, /model and /scan. The same evidence first, approve each step flow applies to the Python and C++ libraries under a simulation or analysis codebase.

Click through it at your own pace, or open it full screen.

The second demo sets up the Safeguard runner on your own machine, approves it with its device code and then scans from the terminal. For a lab that keeps code and data on its own hardware, that is the way to inventory dependencies without moving the repository.

Click through it at your own pace, or open it full screen.

The third demo connects Claude to the Safeguard MCP server with OAuth and lets you choose which tools Claude can use. If you already use a general assistant for research coding, this lets it ask Safeguard which projects carry a given package version, while you control what it can touch.

Click through it at your own pace, or open it full screen.

To try it on your own code, download Safeguard and follow the CLI documentation.

Quick check

Sources

• NobelPrize.org: Press release, Nobel Prize in Chemistry 2024
• NobelPrize.org: Popular information, Nobel Prize in Chemistry 2024
• NobelPrize.org: Press release, Nobel Prize in Physics 2022
• NobelPrize.org: Press release, Nobel Prize in Chemistry 2019
• NobelPrize.org: Fritz Haber facts, Nobel Prize in Chemistry 1918
• Lawrence Livermore National Laboratory: LLNL achieves fusion ignition (13 December 2022)
• US Department of Energy: DOE national laboratory makes history by achieving fusion ignition
• LLNL: Achieving fusion ignition (milestone list)
• NASA Science: Dark matter
• NASA Science: Dark energy
• CERN: Antimatter and the matter antimatter asymmetry problem
• CERN: Extra dimensions, gravitons and tiny black holes
• LHCb collaboration, Observation of charge parity symmetry breaking in baryon decays, Nature (16 July 2025)
• ALPHA collaboration, Observation of the effect of gravity on the motion of antimatter, Nature (27 September 2023)
• DESI collaboration, DESI DR2 Results II: BAO measurements and cosmological constraints, arXiv 2503.14738
• DESI: DR2 results, March 19 guide
• Stanford Encyclopedia of Philosophy: Philosophical issues in quantum theory
• Lee, Kim and Kwon, The first room temperature ambient pressure superconductor, arXiv 2307.12008
• Nature news: LK-99 isn't a superconductor (16 August 2023)
• Nature: Retraction note, Evidence of near ambient superconductivity in a N-doped lutetium hydride (7 November 2023)
• Abramson et al., Accurate structure prediction of biomolecular interactions with AlphaFold 3, Nature (8 May 2024)
• Andersen et al., A rigorous electrochemical ammonia synthesis protocol with quantitative isotope measurements, Nature (22 May 2019)
• Merchant et al., Scaling deep learning for materials discovery, Nature (29 November 2023)
• Google DeepMind: Millions of new materials discovered with deep learning (29 November 2023)
• Szymanski et al., An autonomous laboratory for the accelerated synthesis of inorganic materials, Nature (29 November 2023)
• Nature: Author correction to the A-Lab paper (19 January 2026)
• Leeman et al., Challenges in high throughput inorganic materials prediction and autonomous synthesis, PRX Energy (7 March 2024)
• Degrave et al., Magnetic control of tokamak plasmas through deep reinforcement learning, Nature (16 February 2022)
• Google Quantum AI, Quantum error correction below the surface code threshold, Nature (9 December 2024)
• Reproducible Builds: Definitions
• SLSA: Security levels, version 1.1
• Safeguard: Model benchmarks

Last reviewed: 23 October 2026.