Unveiling the Core of Materials Properties via SISSO and Sensitivity Analysis

Lucas Foppa*, Matthias Scheffler · 2026

Read Unveiling the Core of Materials Properties via SISSO and Sensitivity Analysis by Lucas Foppa*, Matthias Scheffler (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-08122.txt.

Introduction

Foppa and Scheffler use clear AI to uncover physical rules behind matter traits. The SISSO symbolic fitting approach finds rules linking a target trait to a small set of inputs called matter genes, picked from a large pool. Multiple gene mixes can yield equally exact SISSO models, with single genes carrying varied weights, leaving accounts non-unique. The team sets up a slope-based response study that settles this overlap and lifts clarity. The analysis reveals how distinct gene sets encode same information, growing insight.

Catalog record

CREATORLucas Foppa*, Matthias Scheffler
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE27112
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.08122

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Unveiling the Core of Materials Properties via SISSO and Sensitivity Analysis Lucas Foppa*, Matthias Scheffler Abstract Interpretable AI can reveal physical principles governing intricate materials properties by uncovering explicit relationships between physical parameters and target properties. The sure-independence screening and sparsifying operator (SISSO) symbolic-regression approach identifies analytical expressions that correlate a target property with a small set of parameters, termed materials genes , selected from a large pool of candidates. However, multiple gene combinations can yield equally accurate SISSO models, with individual genes contributing with different weights. Here, we establish a derivative-based sensitivity analysis that resolves the non-uniqueness of symbolic-regression descriptions, enhances interpretability, thereby enabling deeper physical insight. This analysis reveals how distinct gene combinations encode equivalent information and identifies valence orbital radii, nuclear charges, and their products as the key quantities governing the equilibrium lattice constant of perovskites. Predictive models linking basic physical parameters to materials properties and functions are key to accelerating materials discovery. Atomistic simulations accurately predict some materials properties. They offer detailed physical insights, but are inappropriate to model properties governed by multiple entangled physical processes. AI and machine learning reveal, based on appropriate data, nonlinear correlations between multiple input parameters, termed primary features, and target properties. 36 ; 38 ; 33 ; 11 These methods might thus capture intricate materials’ properties more effectively than explicit theoretical approaches. However, their flexibility often comes at the cost of interpretability, as many AI models act as “black boxes,” providing limited insight into the physical mechanisms governing the materials properties. To mitigate this problem, primary-feature-importance analyses are often used to identify the most critical primary features for the models, thereby providing more physical insight. Such explainable AI analyses 8 ; 6 can be based on different concepts such permutation of primary features 13 , local approximations such as the local interpretable model-agnostic explanations (LIME) approach 37 , and the SHapley Additive exPlanations (SHAP) method. 24 ; 41 ; 1 As an alternative to black-box models, symbolic regression (SR) 39 ; 43 ; 30 ; 31 ; 46 ; 26 has emerged as an inherently interpretable approach. SR identifies models for materials properties as analytical expressions, thereby rendering explicit a mathematical relationship between the primary features and the target materials property of interest. Some SR approaches can also take into account relationships described by derivatives or integrals. 16 ; 22 The sure independence screening and sparsifying operator (SISSO) method 31 ; 35 has gained prominence due to its deterministic and efficient expression-selection process. 9 ; 10 ; 45 ; 32 ; 42 SISSO begins by generating an immensity of candidate analytical functions from an initial set of physically meaningful primary features which characterize the material and the environment. These functions are formed by iteratively applying (nonlinear) unary and binary mathematical operators such as addition, multiplication, and logarithm, in order to combine the primary features. Then, compressed sensing 14 ; 28 is used to select a small number (often less than 4) of analytical functions that linearly combined by weighting coefficients best correlate with the target property. The SISSO models typically depend only on small number of primary features, selected from the large pool of offered ones. These selected primary features are called materials genes 18 , in analogy to genes in biology and medicine, in order to emphasize their statistical nature and the concept of correlations between these materials genes and the property of interest, as opposed to physical laws. The different genes selected in the SISSO models might impact the property in different extents. Additionally, multiple gene combinations can yield equally accurate SISSO models. Thus, the set of genes required to describe a given materials property is not unique. This hinders deeper physical insights and the decision on what additional data needs to be acquired for accurately modelling the materials property of interest. One strategy used by some authors to obtain SISSO models that depend only on the few most important primary features has been to evaluate the correlations between primary features and exclude primary features that are highly correlated with other primary features before model training. 21 ; 44 However, important correlations, e.g., resulting from the interaction of multiple primary features, i.e., the combination of two primary features according to a binary operator such as difference, might be missed when correlated primary features are excluded prior to model training. Thus, in this paper we emphasize that sensitivity analyses may be preferable to identify the most influential primary features after a model is obtained based on a comprehensive set of primary features. 25 ; 40 ; 3 ; 17 ; 34 Sensitivity analyses examine how changes in an (input) primary feature affect the model target-property description (output). They can provide local sensitivity scores per data point, e.g., per material, or global sensitivity scores averaged over all materials in a dataset. The Sobol method, for instance, is a global sensitivity analysis 40 ; 23 ; 34 that decomposes the variance of the model output into contributions from individual primary features and their interactions. Here, we establish a gradient-based partial-effects (PE) sensitivity analysis to resolve the non-uniqueness of symbolic-regression descriptions and enhance interpretability, enabling deeper physical insight. The PE method 29 ; 4 quantifies the impact of a given primary feature in the model’s output by means of the partial derivative. 4 ; 5 Thus, PE quantifies the weight of a primary feature when the remaining primary features are kept unchanged. PEs provide global and local sensitivity scores and the analysis is less computationally demanding than other widely used ones, as the partial derivatives are obtained analytically. As an example, we demonstrate the power of the PE analysis combined with SISSO for modelling the equilibrium lattice constant ( a_{0} ) of cubic A_{2}BB^{\prime} O 6 double perovskites. Obviously, the concept also works for any other materials property and any other class of materials. It has been also employed for a study in heterogeneous catalysis. 19 In the perovskite formula, we define that B^{\prime} is the more electronegative element than B . Single perovskites with the formula AB O 3 (with B=B^{\prime} ) are also included in the dataset of 4,583 compounds. The target a_{0} was calculated using density functional theory (DFT) with the PBEsol 15 exchange correlation functional and the FHI-aims code. 12 ; 2 As primary features, we offer 23 basic physical parameters. These include properties of free-atoms of the elements A , B and B^{\prime} evaluated with DFT-PBEsol, such as, for example, the radii of s and valence (val) orbitals of the neutral and +1 cation (cat) of free atoms ( r_{s} , r_{\mathrm{val}} , r_{s}^{\mathrm{cat}} , and r_{\mathrm{val}}^{\mathrm{cat}} ), the electron affinity ( EA ), and the ionization potential ( IP ). EA and IP are calculated by the total energy difference between the neutral and charged atoms. The oxidation states of A and the average oxidation state of B and B^{\prime} elements in the perovskite composition ( n_{A} and n_{\bar{B}} ) approximated by integers determined based on the periodic table group of A and charge neutrality of the formula unit, are also included. Note that the charge neutrality condition results in the relation n_{A}+n_{\bar{B}}=6 . Some of the primary features are correlated with each other (see Pearson

Read the complete text