## INTRODUCTION

Fashions within the chemical and bodily sciences have led to each new understanding and new discoveries (*1*) together with new supplies (*2*, *3*). Physics-based fashions span orders of magnitude in size and time, starting from quantum mechanics (*4*) to chemical vegetation (*5*), and naturally seize physics-based constraints (*6*–*8*). Combining fashions throughout scales, referred to as multiscale modeling (*9*), is critical when chemical properties are decided on the quantum stage, however most experiments and related functions exist on the macroscale, equivalent to in heterogeneous catalysis. On the core of mannequin improvement lies the query of accuracy of a physics-based mannequin. Going past sensitivity evaluation (*10*, *11*), there was rising curiosity in quantifying uncertainty, ensuing from correlations in parameters (*12*, *13*) together with different sources of error arising in predicting new supplies (*14*). Along with guaranteeing trustworthiness, error quantification can allow mannequin correctability (*15*, *16*). Nonetheless, uncertainty is an afterthought in precise physics-based mannequin improvement. At present, a mannequin is first constructed deterministically with out systematically accounting for the impact of each modeling errors and lack, or sparseness, of knowledge.

Modeling unsure information has skilled super advances in information science (*17*–*20*); nevertheless, the corresponding fashions are empirical, can fail with out assure, and might violate conservation legal guidelines and constraints. Present approaches for dealing with information based mostly on bodily legal guidelines and chemical concept are, in a way, not really probabilistic and require correlations and causal relationships to be recognized a priori. With the rising dimension of chemistry datasets, it’s virtually unimaginable to use conventional strategies of mannequin improvement to techniques with many sources of interacting error. World sensitivity methods, equivalent to Sobol indices, attribute mannequin variance to mannequin variables and their interplay (hereafter known as “parametric uncertainty”) (*21*). Nevertheless, there are few strategies that work past first-order interactions or quantify the significance of lacking physics or submodels reasonably than parameters (hereafter known as “mannequin uncertainty”) (*22*). Strategies that do exist mannequin lacking physics as stochastic noise that has no construction (*23*, *24*). Due to this fact, there’s a have to develop strategies that each attribute interplay error on to mannequin inputs and supply predictive ensures and, by doing this, to make fashions correctable and finally reliable for predictions and design duties.

Right here, we tackle these points by incorporating error and uncertainty straight into the design of a mannequin. First, we introduce using Bayesian networks (*20*), a category of probabilistic graphical fashions (PGMs) (*25*), frequent in probabilistic synthetic intelligence (AI) (*26*), to combine concurrently and systematically physics- and chemistry-based fashions, information, and professional information. This framework is termed C-PGM (chemistry-PGM). Second, we derive world uncertainty indices that quantify mannequin uncertainty stemming from completely different physics submodels and datasets. This framework generates predictive “worst-case” ensures for Bayesian networks whereas dealing with correlations and causations in heterogeneous information for each parametric and mannequin uncertainties and relies on current work in strong strategies for quantifying probabilistic mannequin uncertainty (*27*, *28*). Our proposed framework, combining AI and uncertainty quantification (UQ), systematically apportions and quantifies uncertainty to create interpretable and correctable fashions; that is carried out via assimilation of knowledge and/or enchancment of bodily fashions to allow reliable AI for chemical sciences. We scale back the complexity of the nondeterministic polynomial time (NP)–onerous downside of studying a PGM by leveraging professional information of the underlying chemistry. We reveal this framework within the prediction of the optimum response charge and oxygen binding power for the oxygen discount response (ORR) utilizing the volcano mannequin. Whereas UQ has been utilized to deterministic volcano-based fashions normally (*29*), and the ORR mannequin particularly (*30*), prior strategies have been restricted each by the physics mannequin’s underlying construction and, importantly, the dearth in interpretability of uncertainty in predictions by way of modeling selections and out there information in several mannequin parts. We reveal that about half of the mannequin uncertainty stems from density practical concept (DFT) uncertainty, comparable error from lack of adequate quantity and high quality of experimental information and from correlations in parameters (~20% every), and the remaining (~10%) from the solvation mannequin. This evaluation supplies a blueprint for prioritizing mannequin parts towards correctability and improved trustworthiness by underscoring the necessity foremost of extra correct digital construction calculations and secondary by higher experiments. We illustrate mannequin correctability with an instance.

## RESULTS

### Physics mannequin for the ORR and the deterministic volcano

Hydrogen gasoline cells can almost double the effectivity of inside combustion engines and depart behind virtually no emissions, particularly if environmentally low footprint H_{2} is offered (*31*). Moreover, the hydrogen gasoline cell is a mature expertise that produces electrical energy through the hydrogen oxidation response on the anode and the ORR on the cathode (Fig. 1B); polymer electrolyte membrane gasoline cells for such a response are commercially out there (*32*). Due to the excessive price of platinum (Pt) catalyst and stability issues of different supplies in an acidic electrolyte, current focus has been on growing alkaline electrolytes. This expertise, whereas extraordinarily promising, ends in slower response charges (by ~2 orders of magnitude in comparison with a Pt/acidic electrolyte) and thus larger units (*33*, *34*). Overcoming slower charges with steady supplies requires discovery of recent, multicomponent catalysts, e.g., core-shell alloys.

The ORR relies on the formation of floor hydroperoxyl (OOH*), from molecular oxygen (O_{2}), and of water (H_{2}O), from floor hydroxide (OH*) (*35*). The whole mechanism (*7*, *36*, *37*) entails 4 electron steps (Fig. 1A) and is described intimately in part S1. Amongst these, reactions R1 and R4 are gradual (*7*). Acceleration of the ORR then interprets into discovering supplies that velocity up the slower of the 2 reactions, R1 and R4. An method to search out new supplies entails era of an exercise mannequin (Fig. 1C) as a perform of descriptor(s) that may be estimated shortly utilizing DFT calculations (*9*). This is named the deterministic volcano (Sabatier’s precept) and has been the important thing mannequin for discovery of recent supplies.

Subsequent, we talk about the human workflow in developing the volcano curve. First, we use a physics equilibrium mannequin to compute the speed *r* from the minimal free power of reactions R1 and R4 (*7*, *38*), such that

(1)the place *okay*_{B} is the Boltzmann fixed and *T* is the temperature. As a substitute of Eq. 1, one may use a extra elaborate mannequin, equivalent to a mean-field microkinetic (detailed response mechanism) or a kinetic Monte Carlo mannequin, which is a extra complicated multiscale mannequin. Such fashions impose conservation legal guidelines (mass conservation and catalyst website steadiness) and are chosen on the premise of professional information. The Gibbs free power Δ*G _{i}* of the ith species is calculated from the digital power (

*E*

_{DFT}) and consists of the zero-point power, temperature results, and an specific solvation power (

*E*

_{solv}) in water, as detailed in Strategies. This calculation entails, once more, physics-based fashions (statistical mechanics right here) and professional information, e.g., in deciding on a solvation mannequin and statistical mechanics fashions. See part S1 for an evidence of the equilibrium mannequin and ensuing system in Eq. 1.

The free energies Δ*G*_{1} and Δ*G*_{4} are computed as linear combos of the free energies of species, whereas accounting for stoichiometry (a constraint), and are regressed versus Δ*G*_{O} (the descriptor); see information in Fig. 1C. Usually, solely two to a few information factors for coinage metals (Ag, Au, and Cu) on the correct leg of the volcano are regressed, particularly if experimental information (as an alternative of DFT information) are used (information similar to dotted traces will not be noticed in most experiments). The intersection of the 2 traces (Fig. 1C) determines the utmost of the volcano curve and supplies optimum materials properties, i.e., the

;

$\mathrm{\Delta}{G}_{\mathrm{O}}^{*}$can then be matched to values of multicomponent supplies to acquire supplies nearer to the tip of the volcano. This “human workflow” (Fig. 1D) supplies a blueprint of the deterministic total mannequin that depends completely on professional information in design and numerous bodily submodels (known as additionally parts) for estimation of a number of key portions.

### Probabilistic AI for chemistry and the probabilistic volcano

Right here, we develop a probabilistic AI-based workflow that augments the human workflow (in Fig. 1D) to create a probabilistic volcano. The mathematical device we use to formulate the probabilistic volcano is the PGM. PGMs signify a studying course of by way of random variables, depicted as vertices of the graphs, which explicitly mannequin their interdependence by way of a graph. This interdependence stems from (i) one variable influencing others, known as causality, depicted by arrows (directed edges); and (ii) correlations amongst variables, depicted by easy (undirected) edges between vertices (see beneath). PGMs are outlined because the parameterized conditional chance distribution (CPD) and for Bayesian networks are outlined such that

$$\begin{array}{c}P(X|\mathrm{\theta})={\displaystyle \prod _{i=1}^{n}}P({X}_{i}|{\mathit{Pa}}_{{X}_{i}},{\mathrm{\theta}}_{{X}_{i}|P{a}_{{X}_{i}}})\text{with}\text{CPD}:P({X}_{i}|{\mathit{Pa}}_{{X}_{i}},\\ {\mathrm{\theta}}_{{X}_{i}|P{a}_{{X}_{i}}}),i=1,\dots ,n\end{array}$$(2)

*Pa _{Xi}* = {

*X*

_{i1}…,

*X*} ⊂ {

_{im}*X*

_{1}…,

*X*} denotes the mother and father of the random variable

_{n}*X*, and θ = {θ

_{i}_{Xi ∣ PaXi}}

^{n}_{i = 1}are the parameters of every CPD,

*P*(

*X*∣

_{i}*Pa*, θ

_{Xi}*). Uppercase “*

_{XiPaXi}*P*” signifies a stochastic mannequin or submodel. A key idea in PGMs is that the random variables are conditionally impartial. This idea is central to developing complicated chance fashions with many parameters and variables, enabling distributed chance computations by “divide and conquer” utilizing graph-theoretic mannequin representations. By combining the conditional chances in Eq. 2, we discover the joint chance distribution of all random variables

*X*. An in depth formalism for the development of the PGM is given in part S4.

Construction studying of graphical fashions is, normally, an NP-hard downside (*39*, *40*) if one considers the combinatorial nature in connecting numerous vertices. We overcome this problem by constraining the directed acyclic graph (DAG) (*41*), representing the probabilistic ORR volcano (Fig. 2A), utilizing area information that features multiscale, multiphysics fashions mentioned above, professional information, and heterogeneous information (experimental and DFT) together with their statistical evaluation.

First, statistical evaluation of knowledge finds hidden correlations or lack thereof between variables and can also be central to constructing the PGM. On this instance, statistical evaluation of the computed formation free power information of O*, OOH*, and OH* signifies correlations amongst information, i.e., connections between vertices (Fig. 2A). Particularly, Δ*G*_{OOH} and Δ*G*_{OH} are correlated with Δ*G*_{O}. The correlation coefficients of Δ*G*_{O} with −Δ*G*_{1} and −Δ*G*_{4} are −0.95 and 0.91, respectively; see part S4 for notes on statistical impartial checks used. Response free energies Δ*G*_{1} and Δ*G*_{4} are linear combos of Δ*G*_{OOH} and Δ*G*_{OH}, respectively; we use the response free energies as dependent vertices, because the response charge relies upon straight on Δ*G*_{1} and Δ*G*_{4}. Subsequently, we select Δ*G*_{O} because the impartial node (descriptor) as a result of, of all of the floor intermediates, it has the fewest levels of freedom (and subsequently native minima) on any given potential power floor for sooner quantum calculations. The collection of the descriptor, which is one other instance of professional information in our C-PGM, establishes causal relationship (route of affect) represented by directed edges from Δ*G*_{O} to Δ*G*_{1} and Δ*G*_{4}. Skilled information can also be leveraged to assign related errors (ω) to vertices and directed edges. Figure 2A (coloured circles) depicts the a number of uncertainties (random variables) ω modeled in every CPD of the PGM and the way these (causally) affect the uncertainty of every vertex. All these causal relationships are modeled by a DAG in Fig. 2A and the Bayesian community in Eq. 3. Causality simplifies the development and UQ evaluation of the PGM. Final, the dearth of an edge between Δ*G*_{1} and Δ*G*_{4} (Fig. 2A) is discovered from conditional independence checks on the DFT information. By eliminating graph edges of uncorrelated components of the graph, the constrained DAG is profoundly easier. A whole, step-by-step dialogue of the construction studying of the ORR C-PGM is included in part S4.1.

The C-PGM construction comprises info from professional information, causalities, physics (bodily fashions, conservation legal guidelines, and different constraints), correlations of knowledge, parameters, and hierarchical priors (priors of a previous) in mannequin studying (*13*, *25*). The bodily which means and estimation of those uncertainties are mentioned beneath. Total, the mannequin for the ORR C-PGM turns into

(3)the place Δ*G*_{O (DFT)} signifies a calculated worth from DFT and all different Δ*G* values signify the “true worth” given errors. Lowercase “*p*” signifies chance densities which can be assumed right here to be Gaussian, thus rendering the C-PGM (Eq. 3) right into a Gaussian Bayesian community (*25*). Notice that this PGM is used as a part of an optimization scheme the place Δ*G*_{O (DFT)} is formulated as a random variable given any worth of the true Δ*G*_{O} and distribution of errors for Δ*G*_{O}. Leveraging the human-based (deterministic) workflow in Fig. 1D, the ORR is modeled as a stochastic optimization downside such that

(4)the place

$\mathrm{\Delta}{G}_{\mathrm{O}}^{*}$ corresponds to the optimum oxygen binding power that maximizes the response charge *r**. It’s handy to compute *okay*_{B}*T* ln (*r**),

(5)

For the remainder of this paper,

$\mathrm{\Delta}{G}_{\mathrm{O}}^{*}$ and *okay*_{B}*T* ln (*r**) are thought-about because the QoIs (portions of curiosity) that have to be optimized.

### Mannequin uncertainty, ensures, nonparametric sensitivity, and contributions to mannequin error for interacting variables

Mannequin uncertainty arises from a number of sources, equivalent to use of sparse information in Fig. 1C, hidden correlations between vertices within the graph, simplified statistics fashions (linear regression between free energies in Fig. 1C and Gaussian approximations of errors; Fig. 2, B to E), and uncertainty in several mannequin parts and variables. These embody errors in experimental information (ω* _{e}*), DFT information (ω

*), solvation energies (ω*

_{d}*), and regressions (correlations) used to find out the optimum*

_{s} (ω* _{c}*); correlation error is accentuated by the small information out there particularly on the correct leg of the volcano. Experimental errors (ω

*) in Δ*

_{e}*G*

_{O}and Δ

*G*

_{OH}will be discovered by repeated measurements in the identical laboratory and between completely different laboratories. Repeated calorimetry and temperature-programmed desorption measurements for the dissociative adsorption enthalpy of O

_{2}in the identical and completely different labs present a distribution of errors for Δ

*G*

_{O}. The distribution of DFT errors (ω

*) is computed by evaluating experimental and calculated (DFT) information throughout numerous metals. The imply worth and SD of errors are offered in desk S1 together with an in depth description of how errors have been calculated in Strategies and part S3. In Fig. 2A, the “father or mother vertex” is set by the route of the arrow such that ω*

_{d}_{e1}is a father or mother of Δ

*G*

_{1}(the kid). These extra uncertainties from a number of sources are proven in Fig. 2 (B to E) and are mixed to construct the PGM mannequin

*P*.

When constructing the C-PGM mannequin *P*, “mannequin uncertainties” come up from the sparsity and high quality of the out there information in several parts of the mannequin, the accuracy of the physics-based submodels, and the information relating to the chance distribution of the errors (Fig. 2, B to E). Consequently, the imply worth of min{ − Δ*G*_{1}, − Δ*G*_{4}} with respect to *P* is itself unsure because the probabilistic mannequin *P* is unsure. For that reason, we take into account *P* as a baseline C-PGM mannequin, i.e., an affordable however inexact “compromise.” Right here, we take *P* to be a Gaussian Bayesian community (see Fig. 2, B to E), the place the error chance distribution perform for every part of the mannequin is approximated as a traditional distribution and is constructed utilizing the nominal datasets and submodels mentioned above. We isolate mannequin uncertainty in every part (CPD) of the whole mannequin (Eq. 3), in distinction to the extra commonplace parametric (aleatoric) uncertainty already included within the stochasticity of *P* itself. We mathematically signify mannequin uncertainty via various (to *P*) fashions *Q* that embody the “true” unknown mannequin *Q**. As examples, fashions *Q* can differ from *P* by (i) changing a number of CPDs in Eq. 3 by extra correct, presumably non-Gaussian CPDs that signify higher the information in Fig. 2 (B to E), (ii) extra correct multiscale physics fashions, and (iii) bigger and extra correct datasets. Quantifying the influence of mannequin uncertainties on predicting the QoI utilizing *P*, as an alternative of higher various fashions *Q*, is mentioned subsequent. Total, growing the mathematical instruments to allow identification of the parts of a PGM that want enchancment is crucial to appropriate the baseline mannequin *P* with minimal sources.

Every mannequin *Q* is related to its personal mannequin misspecification parameter η that quantifies how far another mannequin *Q* is from the baseline mannequin *P* through the Kullback–Leibler (KL) divergence of *Q* to *P, R*(*Q*‖*P*). We use the KL divergence as a result of its chain rule properties that permit us to isolate the influence of particular person mannequin uncertainties of CPDs in Eq. 2 on the QoIs in Fig. 2, see sections S6 and S7. To isolate and rank the influence of every particular person CPD mannequin misspecification (η* _{l}*), we take into account the set of all PGMs

*Q*which can be equivalent to the whole PGM

*P*besides on the

*l*th part CPD (for dependence on the

*l*th mother and father) and fewer than η

*in KL divergence from the baseline CPD*

_{l}*P*(

*X*∣

_{l}*Pa*) whereas sustaining the identical mother and father

_{Xl}*Pa*. We discuss with this household, denoted by

_{Xl}*D*

^{η}

*, because the “ambiguity set” for the*

_{l}*l*th CPD of the PGM

*P*(see part S6 for its mathematical definition). Given the set of PGM’s

*D*

^{η}

*, we develop mannequin uncertainty ensures*

_{l} for the QoI in Eq. 6 as the 2 worst-case eventualities for all doable fashions *Q* in *D*^{η}* _{l}* with respect to the baseline

*P*

(6)

For a given η* _{l}*, the mannequin uncertainty ensures to explain the utmost (worst case) anticipated bias when just one a part of the mannequin within the PGM,

*P*(

*X*∣

_{l}*Pa*), is perturbed inside η

_{Xl}*; subsequently, they measure the influence of mannequin uncertainty in any part (CPD) of the PGM on the QoI. Since η*

_{l}*will not be essentially small, the tactic can also be nonperturbative, i.e., it’s appropriate for each small and huge mannequin perturbations.*

_{l}Equation 6 will be additionally seen as a nonparametric mannequin sensitivity evaluation for PGMs because it entails an infinite dimensional household of mannequin perturbations *D*^{η}* _{l}* of the baseline mannequin

*P.*This household can take into account the sparsity of knowledge by addition of recent or higher-quality information, e.g., higher-level DFT information, various densities to the Gaussians in Fig. 2 (B to E), e.g., richer parametric households or kernel-based CPDs, or extra correct submodels. All these are massive, nonparametric perturbations to the baseline

*P*mannequin. For these causes, Eq. 7 permits one to interpret, reevaluate, and enhance the baseline mannequin by evaluating the contributions of every CPD to the general uncertainty of the QoIs via the (mannequin uncertainty) rating index

(7)

For extra particulars, see theorem 1 in part S6 the place we present that for Gaussian Bayesian networks, the ratios in Eq. 7 are computable.

We will use two methods relating to η* _{l}*. First, η

*can by tuned “by hand” to discover how ranges of uncertainty in every part of the mannequin,*

_{l}*P*(

*X*∣

_{l}*Pa*), have an effect on the QoIs. This method is termed a stress take a look at in analogy to finance the place within the absence of adequate information, fashions are subjected to numerous believable or excessive eventualities. Second, as an alternative of treating η

_{Xl}*as a relentless, we are able to estimate η*

_{l}*because the “distance” between out there information from the unknown actual mannequin and our baseline PGM*

_{l}*P*; we discuss with such η

*’s as information based mostly, in distinction to emphasize checks (see part S8). For instance, the information will be represented by a histogram or a kernel density estimator (KDE) approximation (*

_{l}*42*). On this sense, the contribution to mannequin uncertainty from any error supply is each a perform of its variance and the way distant the error is from the baseline, e.g., the Gaussian CPDs in Fig. 2 (B to E).

Given the error distributions and their Gaussian illustration in PGM mannequin *P*, the anticipated worth of min{ − Δ*G*_{1}, − Δ*G*_{4}} ∣ Δ*G*_{O} (black curve) in Fig. 3A is obtained. The colour bar in Fig. 3A signifies how seemingly a response charge happens with given Δ*G*_{O} for mannequin *P* (aleatoric uncertainty). The grey dashed traces in Fig. 3A correspond to the 2 excessive eventualities (derived in part S5) for all doable fashions *Q* by contemplating uncertainty in all parts. We spotlight in Fig. 3B the anticipated worth (black line) and the extremes (grey dashed traces) when solely the DFT error in Δ*G*_{4} is taken into account. All η values listed below are information based mostly and decide what fashions *Q* are thought-about in building of the bounds; solely PGMs which have a KL divergence that’s lower than or equal to η from the baseline are thought-about. The crimson, orange, and inexperienced traces point out potential QoIs that may be computed; right here, we give attention to the uncertainty within the charge (*y* axis; distinction between crimson traces) and the variability of optimum oxygen binding power (*x* axis; distinction between orange traces) as a proxy of supplies choice; see part S7 for extra particulars.

Utilizing the mannequin uncertainty ensures (Eq. 6), we quantify the uncertainty and its influence on mannequin predictions past the established parametric uncertainty; once more, all η* _{l}*’s are information based mostly. By sourcing the influence of every submodel and/or information, Eq. 7 reveals what information, measurement, and computation needs to be improved. The error within the optimum response charge (Fig. 4A) stems from an almost equal contribution of submodels, particularly by solvation (30%), experiment (18%), DFT (33%), and parameter correlation (18%). The uncertainty within the optimum oxygen free power variability (Fig. 4B), i.e., the supplies prediction, stems from solvation (6%), experiment (8%), DFT (48%), and parameter correlation (37%). Totally different QoIs are influenced to a unique diploma by completely different submodels. In each QoIs, the DFT error stands as essentially the most influential. The correlation between O*, OOH*, and OH* is the subsequent most essential part relating to supplies prediction, whereas solvation is the second ranked part relating to response charge. Such predictions are nonintuitive. Whereas earlier work discovered that parametric-based microscale uncertainties will be dampened in multiscale fashions (

*43*), the outcomes of this work will generalize to any fashions the place fine-scale simulations (equivalent to DFT) are sparse or the macroscale QoIs will be made proportional to the microscale properties. Within the subsequent part, we present that Eq. 6 and the ensuing Fig. 4 will also be deployed to enhance the baseline (purely Gaussian) mannequin

*P*.

### Mannequin correctability enabled by mannequin UQ

Mannequin uncertainty as a result of any submodel or dataset, quantified by η* _{l}* and Eq. 6, will be diminished by choosing a greater submodel or dataset than the unique baseline mannequin

*P*. Clearly, these CPDs that exhibit bigger relative predictive uncertainty in Eq. 6 needs to be prioritized and corrected. In our case examine, lowering the DFT error requires to additional develop DFT practical and strategies, a long-standing pursue not addressed right here. Right here, we illustrate the way to perform such mannequin correctability via an instance that’s possible to do. Particularly, we take into account the mannequin consisting of the information used to assemble the volcano and its statistical illustration as that is the second most influential parameter in supplies prediction. We carried out extra DFT calculations on core-shell bimetallics to create an expanded dataset in comparison with that in Fig. 1C (see Fig. 5A). By doing this, we compute the mannequin sensitivity indices

_{l}for the brand new mannequin utilizing theorem 1 and equation S58. Extra particulars and derivations are included in Strategies.

Figure 5B reveals the discount of mannequin uncertainty ensures, outlined as Eq. 6, that are because of the variance of error and the estimated mannequin misspecification parameter within the correlation between DFT-calculated values of Δ*G*_{4} and Δ*G*_{O}, when extra information (bimetallics) are added. With bimetallic information included, the correlation coefficients of Δ*G*_{O} with −Δ*G*_{1} and −Δ*G*_{4} are −0.95 and 0.95, respectively. The uncertainty is diminished each as a result of improved correlation and diminished SE within the regression coefficients because of extra information.

## METHODS

### DFT calculations

We examine adsorption on the close-packed (111 and 0001) transition metallic surfaces. We choose the bottom power website of O* and OH* for comparability with experiments to find out errors, that are summarized in desk S1. We construct the correlations for bimetallics from the bottom power websites on the (111) and (0001) surfaces of the face-centered (fcc) and hexagonal close-packed (hcp) metals, respectively.

*Vacuum section DFT setup*. We calculated binding energies and vibrational frequencies utilizing the Vienna ab initio Simulation Bundle model 5.4 with the projector-augmented wave technique (*44*). We use the Revised Perdew-Burke-Ernzerhof (RPBE) density practical (*45*) with D3 dispersion corrections (*46*). Simulation strategies embody use of spin-polarized calculations for gas-phase species and ferromagnetic metals, a 3 × 3 × 1 Monkhorst-Pack *okay*-point sampling grid (*47*) for all slab calculations, and a 400-eV airplane wave cutoff. Digital power convergence was set to 10^{−4} eV for the power minimization step and 10^{−6} eV for frequency calculations.

For calculations of gas-phase species, the supercell dimension was 10 × 10 × 10 Å. A Brillouin zone was sampled on the gamma level; a 0.005 eV/Å drive cutoff was utilized in geometry optimizations. For slab calculations, the drive cutoff was set to 0.02 eV/Å with 20 Å of vacuum area. Adsorbate energies have been calculated for OOH*, OH*, and O* on essentially the most steady close-packed floor for fcc and hcp metals. The periodic cell consisted of 4 layers with 16 metallic atoms in every layer; the underside two layers have been fastened at their bulk values, decided utilizing a 15 × 15 × 15 *okay*-point grid with the tetrahedron technique and Blöchl corrections. Bulk metallic lattice constants have been pre-optimized with DFT utilizing the Birch-Murnaghan equation of state (*48*). Zero-point energies are calculated for every adsorbate-surface mixture and for all fuel species. All enter information have been created utilizing the Atomic Simulation Surroundings (*49*).

*Solvation section DFT setup*. We emulate specific solvation calculations from earlier work (*38*) besides that, right here, we fluctuate the variety of water layers. Two to 5 layers of water have been positioned above a Pt(111) floor in a honeycomb sample to simulate the aqueous section above the floor. The double layer of water molecules was discovered to adequately seize water binding energies on Pt(111) and H bonds on the floor (*50*). We decided solvation energies for O* by inserting it in an fcc hole website on the water-covered floor. For OH* and OOH*, solvation energies have been decided by changing a water molecule on the floor with the respective species to find out solvation energies. Apart from the selection of practical, the DFT setup was equivalent to that within the vacuum besides that 9 Pt atoms have been included in every layer to accommodate the honeycomb water construction, the *okay*-point sampling was elevated to 4 × 4 × 1, and the airplane wave cutoff was elevated to 450 eV. To supply preliminary geometries, the Perdew-Burke-Ernzerhof (PBE) practical (*51*) was used for all solvation calculations. Solvation power calculations on Pt(111) utilizing the PBE practical don’t trigger inconsistencies with using the RPBE practical for vacuum section calculations. Granda-Marulanda *et al*. (*52*) confirmed that on a number of 111 and 0001 surfaces, the typical distinction in OH* solvation utilizing the PBE and RPBE functionals with dispersion corrections was 0.03 eV; the SDs utilizing these functionals have been comparable at 0.08 and 0.11 eV, respectively. Quite than modifications in solvation throughout completely different surfaces, we examine the variance in solvation power related to the variety of specific water layers used. As a result of energies from PBE and RPBE are correlated, the variance in solvation power with respect to variety of water layers is anticipated to be comparable.

*Temperature results*. Temperature results at 298 Okay have been calculated utilizing statistical thermodynamics together with the harmonic and very best fuel approximations (*53*). Each warmth capability and entropy results have been included in calculating Gibbs free energies used within the volcano curves. Entropy was eliminated when evaluating to experimental enthalpies as mentioned in part S3.

### Deriving mannequin sensitivity indices

Utilizing strong and scalable UQ strategies for common probabilistic fashions (*27*, *28*, *54*) as a place to begin, we outline “ambiguity units” round a baseline mannequin *P* and “predictive uncertainty for QoIs.” Though the definitions of predictive uncertainty (part S6) and mannequin sensitivity indices (Eq. 6) are pure and reasonably intuitive, it’s not apparent that they’re virtually computable. A key mathematical discovering for PGMs, demonstrated in theorem S1, is that that the ensures

will be computed precisely utilizing a variational system for the KL divergence and the chain rule for the KL divergence; the latter level additionally justifies using the KL divergence in defining the nonparametric formulation of the mannequin sensitivity indices. Within the case the place *P* is a Gaussian Bayesian community (*G*), the rating indices in Eq. 7 are computed utilizing Eq. 9.

### Deciding on new high-quality information or improved bodily mannequin for mannequin correctability

Given a baseline mannequin *P* and the sparse dataset for every submodel sampled from an unknown mannequin *Q*, we are able to construct an improved baseline mannequin *P* for our ORR mannequin following the steps beneath.

Step 1: Discover appropriate data-based η* _{l}*’s:

(8)the place *Q* is the surrogate mannequin given by the KDE/histogram, utilizing eqs. S94 and S95.

Step 2: Calculate the mannequin uncertainty ensures for a given QoI utilizing eq. S58 (or eq. S60 for the final PGM)

$${J}_{l}^{\pm}(\text{QoI},P;{\mathrm{\eta}}_{l})\text{for}\text{all}\text{PGM}\text{vertices}l$$Step 3: Choose the *l** part *X*_{l}* of the PGM with the worst ensures

(highest values).

Step 4: Scale back

${J}_{{l}^{*}}^{+}(\text{QoI},P;{\mathrm{\eta}}_{{l}^{*}})$ based mostly on eq. S58. For QoI(*X*) = min { − Δ*G*_{1}, − Δ*G*_{4}} ∣ Δ*G*_{O}, we’ve got that

(9)the place

$\overline{{F}_{{l}^{*}}}({\mathrm{\omega}}_{{l}^{*}})={\mathbb{E}}_{{P}_{{\{{\mathrm{\omega}}_{{l}^{*}}\}}^{c}}}[\text{QoI}]-{\mathbb{E}}_{P}[\text{QoI}]$. For the *l** part(s) of the PGM, we search essentially the most helpful extra information, particularly the information that tightens (reduces) the ensures in Eq. 9. Notice that the ensures include two components: the second producing perform (MGF) and the mannequin misspecification parameter η. Due to this fact, including informative information can scale back the MGF in Eq. 9 (and, thus, the uncertainty ensures

); because the MGF consists of all moments, and, particularly, the variance (*27*), extra information can enhance mannequin *P* and scale back the mannequin misspecification η (see part S8).

**Acknowledgements: **J.L.L. and D.G.V. acknowledge helpful discussions with J. Lym, Okay. Alexopoulos, and G. Wittreich. **Funding:** The analysis of M.A.Okay. was partially supported by NSF TRIPODS CISE-1934846 and the Air Pressure Workplace of Scientific Analysis (AFOSR) beneath the grant FA-9550-18-1-0214. The analysis of J.F. was partially supported by the Protection Superior Analysis Initiatives Company (DARPA) EQUiPS program beneath the grant W911NF1520122, and a part of J.F.’s work was accomplished throughout his PhD at UMass Amherst. J.L.L. and D.G.V. acknowledge assist by the RAPID Manufacturing Institute, supported by the Division of Vitality (DOE) Superior Manufacturing Workplace (AMO), award quantity DE-EE0007888-9.5. RAPID tasks on the College of Delaware are additionally made doable, partly, by funding offered by the State of Delaware. The Delaware Vitality Institute acknowledges the assist and partnership of the State of Delaware in furthering the important scientific analysis being carried out via the RAPID tasks. J.L.L.’s analysis used sources of the Nationwide Vitality Analysis Scientific Computing Heart (NERSC), a U.S. Division of Vitality Workplace of Science Person Facility operated beneath contract no. DE-AC02-05CH11231. The 2019 to 2020 Blue Waters Graduate Fellowship to J.L.L. can also be acknowledged. **Creator contributions:** J.F. developed and carried out the PGM mannequin. J.F. and M.A.Okay. developed UQ for PGMs. M.A.Okay. conceived using PGMs for mannequin uncertainty, in addition to the associated information-theoretic instruments. J.L.L. developed the bodily mannequin, and D.G.V. considered the concept to make UQ explainable, to use the PGM mannequin to the ORR, and the necessity to apportion error to completely different mannequin inputs for sparse information and lacking physics. All authors contributed to writing. **Competing pursuits:** The authors declare that they don’t have any competing pursuits. **Information and supplies availability:** All information wanted to judge the conclusions within the paper are current within the paper and/or the Supplementary Supplies. All underlying DFT and experimental information can be found on Zenodo. Software program might be made out there upon request. Extra information associated to this paper could also be requested from the authors.