When LLMs Learn Physics: Taming Symbolic Regression in Materials Science

Formula discovery sounds like the part of science where artificial intelligence should behave like a heroic mathematician: stare at data, discover a law, and write down a clean equation while everyone else politely applauds.

That is the cinematic version. The actual engineering problem is less glamorous and much more useful.

Symbolic regression already searches for equations. Given enough variables, operators, constants, and patience, it can produce formulas that fit data. The trouble is that “fits data” and “means something physically” are not the same sentence. In a high-dimensional materials dataset, symbolic regression can wander through a forest of plausible-looking algebra and return a formula that is accurate, ornate, and scientifically suspicious. A spreadsheet can also produce a trendline. We do not usually call that physics.

The paper behind today’s article, Discovery of Interpretable Physical Laws in Materials via Language-Model-Guided Symbolic Regression, proposes LangLaw, a hybrid framework that uses a large language model to guide symbolic regression rather than replace it.¹ That distinction matters. The LLM does not act as a lone genius generating laws from raw numbers. It acts more like a scientifically informed search manager: it reads feature descriptions, suggests physically relevant variables, sets search constraints, reviews earlier symbolic regression attempts, and helps narrow the next search round.

That is a less romantic role for the LLM. It is also a more believable one.

The real bottleneck is not regression. It is the search space.

Symbolic regression is attractive because it promises explicit formulas. Instead of training a black-box predictor for a material property, the researcher can obtain an equation linking the property to physical descriptors: electronegativity, ionic radius, valence electrons, lattice parameters, oxidation numbers, and so on.

That is the dream. The nightmare is combinatorial explosion.

Suppose the model has many candidate features and a small set of mathematical operators: addition, subtraction, multiplication, division, square roots, logarithms, exponentiation. Even before we discuss constants and nesting depth, the number of possible formulas grows violently. Traditional symbolic regression therefore faces an ugly trade-off:

Materials science makes this harder because many datasets are small. The paper studies three tasks: perovskite bulk modulus, lead-free double perovskite band gap, and oxygen evolution reaction activity. These are not internet-scale datasets. They are closer to the real R&D world, where every data point may require expensive computation, synthesis, measurement, or curation.

Deep learning is often excellent when data are abundant. In small-data scientific discovery, it may become a very expensive way to overfit politely.

LangLaw’s premise is that the LLM should not compete with symbolic regression in numerical search. Symbolic regression is already designed for that. The LLM should instead reduce the amount of nonsense the regression engine is asked to consider.

That is the central mechanism.

LangLaw gives the LLM the job it can plausibly do

The paper’s workflow is an iterative loop:

1. The LLM reads the dataset description and the meanings of the input features.
2. It proposes a constrained symbolic regression search: selected variables, tree depth, and evolutionary iteration settings.
3. PySR performs the actual symbolic regression.
4. Candidate formulas are evaluated by error and complexity.
5. The results are stored in an Experience Pool.
6. The LLM reviews previous attempts and refines the next round.

This is not “LLM discovers physics.” It is “LLM helps symbolic regression stop looking under every mathematically legal rock.” Less magical, more operational. Usually a good sign.

The paper reports that this guidance reduces the effective search space by approximately $10^5$. That number should not be read as a universal law of AI-assisted science. It is evidence that, in their setup, language-guided feature and parameter selection materially narrows the search. The important business translation is not the exact multiplier. It is the architecture: use language models to encode domain priors into search constraints, then let specialized numerical tools do the numerical work.

A useful way to view LangLaw is as a division-of-labor design:

That table is probably the most important part of the paper for business readers. Many AI workflows fail because they assign the wrong task to the model. LangLaw is interesting because it gives the LLM a task that resembles scientific reasoning and search design, not blind numerical optimization.

Bulk modulus: the strongest evidence because simplicity survives out-of-distribution testing

The first test concerns the bulk modulus $B_0$ of perovskites, a measure of resistance to uniform compression. For materials design, bulk modulus is connected to mechanical stability. A formula that predicts it while remaining interpretable is more valuable than a black-box prediction, because it can guide reasoning about which chemical features matter.

The paper compares LangLaw with prior empirical and symbolic-regression formulas. A previous HI-SISSO formula uses multiple coupled terms and fitted exponents. LangLaw finds a simpler expression:

Here, $EA_B$ is the electron affinity of the B-site ion, $IP_B$ is its ionization potential, $n_A$ is the oxidation number of the A-site ion, $a_0$ is the lattice parameter, and $EN_B$ is the electronegativity of the B-site atom.

The structure matters more than the algebraic neatness. The first term, $-\frac{EA_B}{IP_B}$, can be read as a proxy for electron-cloud softness. Higher electron affinity and lower ionization potential imply a more polarizable cloud, which tends to reduce resistance to compression. The second term combines a lattice and charge-related component with an electronegativity correction. The authors interpret it as an effective linear proxy for a nonlinear term in the earlier Verma-Kumar formula, with $EN_B$ adjusting for ionic bond weakening.

This is where LangLaw becomes more than curve fitting. The discovered formula is not just shorter; it can be discussed in the language of bonding, polarizability, and lattice stiffness.

The stronger test is out-of-distribution generalization. The authors test the formula on 10 perovskites selected from 7,308 single and double perovskite structures. These are treated as OOD because their bulk modulus values and double perovskite structures are rare in the training set. LangLaw achieves an OOD RMSE of 0.0851 eV/Å$^3$, compared with 0.411 for HI-SISSO, 0.167 for ALIGNN, and 0.401 for CGCNN.

That result does not prove LangLaw will generalize to every materials family. It does show something more specific and more useful: in this case, the simpler physically guided formula transferred better than more complex symbolic regression and deep learning baselines.

For business R&D, this is the part worth underlining. In early-stage materials screening, the costliest failure is not only bad prediction. It is false confidence from a model that fits the known dataset and collapses outside it. A compact formula that transfers better to unusual candidates is operationally valuable because unusual candidates are often exactly what R&D teams are trying to find.

Band gap: LangLaw matches accuracy but wins mainly on concision

The second task is band gap prediction for lead-free double perovskites, $A_2BB’X_6$. Band gap is central to screening optoelectronic and photovoltaic materials. The dataset contains 745 materials.

LangLaw discovers this formula:

where $V_A$ and $V_B$ are valence electrons of the A- and B-site atoms, $R_X$ is the ionic radius of the X-site anion, and $X_X$ and $X_{B’}$ are electronegativities.

The paper compares this with a SISSO-derived formula. Interestingly, the two formulas share a similar core term involving $\frac{X_X^3}{V_B^4}$, suggesting that both methods detect a similar controlling relationship. They also contain comparable denominator structures involving $X_{B’}$ and $V_A$, though the exact terms differ. The authors note that one extra variable in the LangLaw expression, $R_X$, varies only slightly in the dataset; replacing it with the mean produces only a slight prediction-error increase.

That observation has two readings.

The positive reading: LangLaw rediscovers a physically meaningful structure close to SISSO while producing a more concise expression. This supports the idea that language-guided search can reach scientifically sensible formulas.

The cautious reading: if one term varies little across the dataset, the formula’s apparent sophistication may partly reflect dataset-specific regularity. That does not make it useless. It simply means the formula should be treated as a screening relation under similar chemical-space assumptions, not as a universal band gap law handed down from the mountain.

In Table 1, LangLaw and SISSO both report band gap RMSE of 0.672 eV, while LangLaw uses lower formula complexity: 19 versus 39. LLM-SR reports a similar RMSE of 0.669 but much higher complexity, 70. CGCNN and ALIGNN perform worse on this task, with RMSE values of 1.053 and 1.114.

The business message is therefore more restrained than for bulk modulus. LangLaw does not clearly beat SISSO on band gap error. It matches the reported accuracy while producing a simpler formula. In scientific workflows, that can still matter. Simpler formulas are easier to audit, easier to communicate, and easier to use as hypothesis generators.

Accuracy is useful. Interpretability is useful. The rare thing is getting both without turning the equation into decorative plumbing.

OER activity: small data makes the mechanism more relevant than the leaderboard

The third task concerns oxygen evolution reaction activity in oxide perovskites. OER is a rate-limiting step in electrocatalysis, so better screening relations can help identify promising catalyst candidates. The dataset is extremely small: 18 synthesized ABO$_3$ perovskites measured under comparable experimental conditions, plus a prior test set of 5 high-activity stable perovskites screened from 3,545 candidates.

The previous GPSR formula is:

where $\mu$ is the octahedral factor and $t$ is the tolerance factor.

LangLaw discovers:

The paper notes that the coefficient attached to the $t$ term is very small. Numerically, the term involving $t$ contributes around 0.1, much smaller than the constant 3.24. The authors then examine formulas containing only $\mu$ and find that their prediction errors remain lower than the GPSR formula.

This part of the paper is best read as an interpretability exercise, not a grand claim about catalytic discovery. With only 18 training points, no serious reader should pretend that the model has conquered OER chemistry. What the result does suggest is that LangLaw can recover a compact relationship centered on known geometric descriptors, and that the octahedral factor $\mu$ may dominate within this dataset.

Table 1 reports that for in-distribution OER, LLM-SR achieves the lowest MAE at 0.0147 eV, while LangLaw reports 0.0187 and GPSR 0.0253. For OOD OER, GPSR reports 0.0209, LangLaw 0.0225, and LLM-SR 0.108.

This is an important nuance. LangLaw is not the best on every metric. It is better than GPSR in in-distribution error and far better than LLM-SR on OOD error, but GPSR slightly beats it on the reported OOD OER MAE. A sloppy summary would say “LangLaw wins.” The paper’s actual evidence is more interesting: LangLaw often gives a favorable balance of simplicity, interpretability, and transferability, but the balance depends on the property and benchmark.

That is not a weakness in the article. That is where the article stops behaving like a brochure.

What the experiments actually support

The paper’s evidence is easier to interpret if we separate main evidence, comparisons, and boundary tests.

This distinction matters because the paper sits in a crowded and noisy conversation: “Can LLMs do science?” That question is badly posed. Science is not one task.

LangLaw answers a narrower question: can an LLM’s scientific prior knowledge guide a symbolic regression engine toward simpler, more physically coherent formulas in small-data materials problems? The paper’s answer is: yes, with promising evidence across three cases, and strongest evidence in the bulk modulus experiment.

The business value is not automatic discovery. It is cheaper hypothesis filtering.

For R&D-heavy businesses, the practical value of LangLaw is easy to exaggerate and easy to miss.

The exaggerated version says: “LLMs will discover new laws of materials science.” Perhaps someday. But that is not what this paper proves.

The more useful version says: “LLMs can help scientific search tools spend less time exploring physically silly regions of the hypothesis space.” That sounds modest. It is not.

In industrial R&D, a large amount of time is spent deciding which candidates deserve more expensive evaluation. The direct cost may be computation, lab synthesis, measurement, or expert review. The indirect cost is attention. A model that returns an inscrutable prediction may help ranking. A model that returns a compact formula can help discussion: why this candidate, which descriptor matters, what trade-off appears to be driving the result, and whether the relation makes physical sense.

LangLaw is therefore closer to decision support than autonomous discovery. Its business value is not replacing materials scientists. It is improving the throughput and quality of their hypothesis-generation loop.

A realistic workflow might look like this:

Notice what is missing: “push button, get truth.” Good. That product category should stay missing.

For businesses in batteries, catalysts, semiconductors, photovoltaics, alloys, and specialty chemicals, the likely near-term value is not a fully autonomous scientist. It is a better scientific assistant that proposes constrained searches, records what has been tried, and produces formulas that experts can argue with. In R&D, an argument-worthy formula is often more useful than a black-box score with three decimal places.

Why the LLM should guide, not replace, the equation search

The paper explicitly contrasts LangLaw with direct LLM-based symbolic regression, including LLM-SR. This is a critical design lesson.

LLMs are trained on text and code. They can reason over scientific concepts, retrieve patterns from pretraining, explain variable meanings, and produce plausible mathematical forms. But high-dimensional numerical fitting is not their native strength. Asking an LLM to directly discover equations from complex data is like asking a corporate strategist to personally operate a CNC machine because both activities involve “precision.” Technically possible in a demo. Not a reliable operating model.

LangLaw avoids that trap. The LLM handles semantic and heuristic guidance. PySR handles evolutionary formula search and constant optimization. The Experience Pool gives the LLM memory of prior attempts, so the next search is not independent of the last one.

This is a generalizable AI architecture pattern:

Use the LLM to structure the search; use specialized tools to execute the search; use memory to make the next search less stupid.

That pattern is relevant beyond materials science. It applies to process optimization, quantitative finance research, drug candidate screening, industrial fault diagnosis, and any domain where teams need interpretable hypotheses from limited data.

The key is that the LLM’s output must be operationally constrained. It should not merely “suggest ideas.” It should produce search parameters, feature subsets, candidate reasoning, and revision instructions that downstream tools can actually use.

Otherwise, we are back to AI-flavored brainstorming, the enterprise software equivalent of scented candles.

Boundaries: where the paper is strong, and where buyers should not overread it

The paper’s strongest contribution is architectural. It shows a credible division of labor between scientific language reasoning and symbolic numerical search. It also provides evidence that this design can produce compact and interpretable formulas across three materials-property tasks.

The limitations are equally important.

First, the evidence covers three tasks, all within perovskite-related materials contexts. That is a meaningful testbed, not a universal proof across materials science.

Second, formula validity remains domain-dependent. A symbolic expression can be interpretable and still be wrong outside its chemical regime. The band gap case is a useful reminder: some terms may behave like near-constants within the dataset. That can still be valuable for screening, but it narrows the scope of interpretation.

Third, LLM guidance depends on feature descriptions. If the feature set is poorly designed, mislabeled, or missing key physical descriptors, the LLM cannot magically recover what is absent. It may only prune the wrong forest more efficiently.

Fourth, the framework still requires scientific validation. The output formulas should be treated as candidates: inspect them, test them, compare them with known mechanisms, and evaluate them under distribution shift. In other words, do science. Tragic, but necessary.

Finally, the benchmark comparisons are not uniform victories. LangLaw performs especially well on bulk modulus, including OOD generalization. On band gap, it matches SISSO’s reported RMSE with lower complexity. On OER, it improves over GPSR in-distribution but is slightly behind GPSR on the reported OOD MAE, while doing much better than LLM-SR on OOD. The conclusion is not “LangLaw wins every leaderboard.” The conclusion is “LLM-guided symbolic regression is a serious design pattern for interpretable scientific modeling.”

That conclusion is stronger because it is narrower.

The bigger lesson: AI for science needs tool discipline

LangLaw’s most useful lesson is not that LLMs “learn physics.” The title can have a little drama; the architecture should not.

The real lesson is tool discipline. The paper does not ask the LLM to become a universal regression engine. It gives the model a role aligned with what language models can plausibly contribute: reading scientific descriptions, applying prior knowledge, rejecting implausible feature combinations, learning from earlier attempts, and steering a specialized optimizer.

That is the kind of AI system design Cognaptus likes: not mystical, not decorative, and not pretending that every workflow is a chatbot waiting to happen.

For materials R&D, the practical promise is a faster and more interpretable hypothesis loop. Scientists still choose descriptors, inspect equations, test mechanisms, and decide what deserves experimental cost. LangLaw-style systems may reduce the number of bad searches they need to endure before reaching a useful formula.

That is not a small thing. In scientific discovery, progress often comes from searching the right smaller space.

And if an LLM can help make the search space smaller without pretending to be Newton, that is already a good day at the lab.

Cognaptus: Automate the Present, Incubate the Future.