Cosmology has entered a precision era in which discrepancies between independent datasets, most notably the $H_0$ and $S_8$ tensions, have become robust and statistically significant. These tensions are no longer isolated anomalies but increasingly appear as global consistency constraints on the underlying cosmological model, defining what we will refer to here as a \emph{consistency triangle} of background expansion ($H_0$), structure-growth amplitude ($S_8$), and the redshift-dependence of growth - summarised by the growth index $γ$, or equivalently the shape of $fσ_8(z)$. The third vertex is non-trivial because in modified-gravity scenarios with a redshift-dependent effective gravitational coupling, growth amplitude and growth shape evolve independently, breaking the rigid coupling characteristic of $Λ$CDM. In this work, we use $f(Q)$ gravity as a test case for this emerging paradigm. By drawing on a focused set of recent Bayesian and dynamical-system analyses of the three best-studied functional families - power-law, exponential, and logarithmic - we show that while $f(Q)$ models can alleviate individual tensions, the requirement of simultaneous consistency across $H_0$, $S_8$ and the growth index severely restricts the viable parameter space. A bulk-viscous extension is then briefly examined as a representative illustration of how additional matter-sector freedom is constrained by the same consistency requirement. Our reading of the current literature supports the view that cosmological tensions should be interpreted as global consistency conditions, and that viable extensions of $Λ$CDM must satisfy this multi-probe constraint \cite{CosmoVerse,DiValentino2025Corfu}. Within this framework, only a restricted subset of $f(Q)$ models remains competitive.

We study the design of functional incentive mechanisms for dynamical systems, in which a leader designs a fixed incentive function to motivate a self-interested follower to actuate the system beneficially over an extended horizon, without real-time revision of the incentive. This stands in contrast to the adaptive paradigm, in which the incentive is itself a continuously updated control variable. We formalize the problem as a discrete-time bi-level optimal control problem and derive analytical results for the linear-quadratic case with bilinear incentives and a myopic follower. Specifically, we establish a necessary and sufficient stability condition for the induced closed-loop system, derive a closed-form expression for the gradient of the expected leader cost with respect to the incentive parameter matrix, and obtain a fully closed-form cost expression in the scalar setting. Based on the latter, explicit characterizations of the optimal incentive parameter are provided in two asymptotic regimes: the infinite-horizon limit and the limit of high follower cost. For long horizons, the optimal incentive is shown to become independent of the follower's private cost parameter, with direct implications for robust mechanism design under private information.

We study theoretically the spin-orbit interaction of electrons confined in a tubular semiconductor nanowire, between an inner semiconductor core and an outer metallic extra shell. A band off-offset potential is present at the inner semiconductor-semiconductor interface and a more complex potential barrier at the outer metal-semiconductor contact. The cross section of the nanowire has a hexagonal geometry. We use a model derived from the k-dot-p method, and discuss the effects of the interface potentials on the strength of the spin-orbit coupling and on the localization of the wave functions within the semiconductor shell

In this paper, we extend our method [1] for FMCW radar mutual interference mitigation (IM) based on the discrete fractional Fourier transform (DFrFT). Firstly, we propose a radar signal processing chain including our DFrFT-based IM for real-valued receivers, which we compare to reference algorithms on a synthetic data set. We then reduce computational complexity by reformulating DFrFT-based IM in terms of sparse update signals, which enables mitigation of multiple interferences simultaneously. Finally, we conduct a case study on measurement data and show that our method is compatible with real-world environments.

Population protocols are a model of distributed computation in which a collection of indistinguishable finite-state agents interact randomly in pairs to decide a predicate of their initial configuration. The agents decide by achieving a stable consensus on whether the predicate holds or not. It is known that population protocols can decide exactly the predicates expressible in Presburger arithmetic. Recently, Lossin et al. have introduced a notion of protocol robustness against adversarial crash failures. They show that all atomic Presburger predicates can be decided by robust protocols, and ask whether the same holds for every Presburger predicate. We make progress towards settling this question by proving that all predicates expressible in monadic Presburger arithmetic have robust protocols. In addition, we analyze the cost of robustness in terms of state complexity. We study the ratio between the number of states of the smallest robust protocol for a given predicate and the smallest protocol for it. We show that the cost of robustness is at least double exponential in the size of the predicate, and prove that the robust protocols by Lossin et al. for threshold predicates x >= k have optimal state complexity.

We consider the problem of finding the (unique) minimal Walrasian equilibrium price in multi-item, multi-unit auction models: there are multiple indivisible items for sale, with several units of each item, and a bidder may be interested in buying more than one copy of each item. In its special case with unit-demand bidders, where each bidder demands at most one unit of any item, Andersson, Andersson, and Talman (2013) proposed a general framework of ascending auction algorithms based on the concept of excess-demand item set. This paper extends this approach to the multi-unit case by exploiting the discrete convexity of the Lyapunov function associated with the auction model. In particular, we make use of the facts that (i) the equilibrium price vectors are characterized as the minimizers of the Lyapunov function, (ii) the Lyapunov function is an instance of an L-natural-convex function, and (iii) a concept generalizing ``excess-demand item set'' can be defined in L-natural-convex function minimization in general.

In this paper, we compute the iterated Aluthge transforms $\widetilde{C_ϕ}^{(n)}$ of the composition operator $C_ϕ$ on the weighted Bergman spaces $\mathcal{A}_α^2(\mathbb{D})$, where $ϕ(z)=az+(1-a)$ for $0<a<1$. Also, we obtain the norm and numerical radius of $\widetilde{C_ϕ}^{(n)}$ on $\mathcal{A}_α^2(\mathbb{D})$. We establish that $\widetilde{C_ϕ}^{(n)}$ converges in the strong operator topology on $\mathcal{A}_α^2(\mathbb{D})$. The purpose of this paper is to examine the results of \cite{jung2015iterated} for the weighted Bergman spaces $\mathcal{A}_α^2(\mathbb{D})$. Additionally, by using the iterated Aluthge transforms of $C_ϕ^*$ on $\mathcal{A}_α^2(\mathbb{D})$, we derive the iterated Aluthge transforms of $C_σ$, where $\displaystyleσ(z)=\frac{az}{-(1-a)z+1}$ for $0<a<1$, on some weighted Hardy space $H^2(β_α)$ and study its convergence. Finally, we raise some questions that emerge from these findings.

Photon counting is a cornerstone of quantum optics. Here, we demonstrate precisely counting from 0 to over 9000 photons, beating the Poisson noise limit by at least $4.1~\mathrm{dB}$ across this range. We achieve sub-single-photon precision up to 276 photons per pulse. To do so, we multiplex eight intrinsically photon-number-resolving superconducting nanowire single-photon detectors across 128 temporal modes. We use a model-informed characterization of each of the 1024 detection bins, for optimal precision. We perform quantum detector tomography to reconstruct the positive operator valued measures (POVMs) of the complete device, which consists of $1.38\cdot10^8$ matrix elements. At the repetition rate of our experiment of $80~\mathrm{kHz}$, we can precisely count photons corresponding to an optical power of approximately $71~\mathrm{pW}$, bridging the gap from single-photon measurements to high-sensitivity optical power meters. A photon-number-resolving detector of this size, and the tools used to analyze it, will become increasingly important to characterize large quantum states, as well as tasks in precision metrology and optical power standards.

Toxic cyanobacterial blooms are a growing environmental concern that affects freshwater ecosystems, drinking water supplies, and public health. The cyanobacterium Microcystis is among the most important bloom forming species. It often grows in large colonies, which enhances its flotation, reduces grazing, and improves nutrient regulation. Microcystis cells are held together by a matrix of extracellular polymeric substances (EPS), making colony mechanics crucial for bloom formation. However, an analysis of the biomechanical properties of cyanobacterial colonies, and how these properties relate to environmental conditions like nutrient availability, remains largely missing. Here, we use micropipette force sensors to quantify the linear and non-linear mechanical properties of individual colonies at single-cell resolution. Bulk shear rheology complements these measurements by probing macroscopic properties. The measured tensile strength and yield stress are broadly comparable to those of bacterial biofilms and are far greater than the hydrodynamic stresses typically found in wind-mixed lakes. This implies that cyanobacterial colonies are highly resistant to fragmentation by natural mixing processes. We also show that low nutrient availability, particularly low phosphorus, produced stronger colonies, suggesting structural changes in the EPS. Overall, our results establish mechanical testing as a tool for a more complete and physically grounded understanding of cyanobacterial colony formation.

This work studies numerical integration by the Möbius-transformed trapezoidal rule, which combines the classical trapezoidal rule with a change of variables induced by a Möbius transformation that maps the unit circle onto the real line. It is shown that this method achieves the optimal convergence rate for a polynomially weighted integral over the real line if the integrand lives in a related polynomially weighted Sobolev space with positive integer smoothness index. This result can also be generalized in a slightly weaker form for fractional smoothness indices via complex interpolation of function spaces. The algorithm only requires pointwise evaluations of the weight and the target integrand at prescribed nodes that do not depend on the integrand and weight in question. The established theoretical convergence rates are verified by numerical experiments.

We introduce the TemporallyEdgeDisjointScheduleCompletion (TEDSC) problem in which we need to cover a set of temporal edge demands $D$ by routing $k$ temporal walks through a directed static graph while remaining temporally edge disjoint. This problem combines the temporal aspects of train routing and passenger demands with the static nature of real-world rail networks. We present a polynomial time algorithm for TEDSC. Motivated by real world constraints, we next investigate two restricted variants of TEDSC in which each walk can only travel for some bounded distance or time $h$. We show that both are tractable when parameterized by $k + h$, but hard for $h$ and $k + |D|$. If we fix the underlying network, the two problems exhibit distinct complexities: The distance variant remains $W[1]$-hard parameterized by $k$ even on a path of three vertices whereas the time variant admits an FPT algorithm on any fixed star. Finally, we show how to approximate the number of required walks up to a factor of $(2-h^{-1})$.

Multispectral polarisation imaging has a broad range of applications, from biological cell imaging to agricultural remote surveying. For such applications, especially involving lightweight unmanned aerial vehicles like drones, it is necessary to have compact, single-shot, efficient optical systems. We present a metasurface design that diffractively separates a scene into spectral and polarimetric measurements with a single optical component, operating for 532 nm and 700 nm in a single-shot imaging system. The polarisation imaging performance of the design is shown to be robust to both spectral and angular bandwidths, and multispectral polarimetry is demonstrated experimentally.

Open-source projects often rely on a small group of highly active contributors known as hero developers. Prior work shows that hero developers are common in many OSS and enterprise projects, yet who qualifies as a hero depends heavily on the chosen contribution metric. Code-based metrics identify implementation-focused developers, whereas discussion-based metrics highlight coordination and communication; these metrics capture distinct facets of contribution. We conducted a measurement-sensitive study of multifaceted heroism across 77 Apache Software Foundation projects using three technical measures (commit count, distinct files touched, churn) and two social measures (issue-comment count, number of distinct issues commented on). We examined hero prevalence, overlap among hero sets, and severity-wise bug-fixing outcomes via fix and reopen rates. Results show that hero projects are common under all measures, but identified heroes differ substantially across facets. The pooled Jaccard overlap between technical and social hero sets is only 0.10. Cross-facet asymmetry is evident: 71.4% of technical heroes exhibit strong social activity, while only 24.2% of social heroes show strong technical activity. Fix-rate and reopen-rate differences are modest, yet hero-category rankings vary across severity levels and outcome measures. These findings indicate that heroism is not a single, metric-independent role. A multifaceted perspective offers a more reliable understanding of key contributors and better supports developer prioritisation and severity-aware bug assignment.

Measurements at LHC show an increased production of strange hadrons with charged multiplicity in pp collisions, which is not described by the Lund String Model (with the Monash tune) implemented in PYTHIA. This work investigates string closepacking, a mechanism invoked during hadronization where overlapping strings create a background field that increases the effective string tension. This reduces strangeness suppression, effectively enhancing production. The model also incorporates an option for "popcorn destructive interference", which suppresses baryon production, to address the non-strange $p/π$ ratio, utilizing color algebra arguments; and an option for "strange junctions", which enhances strangeness specifically within the baryon sector. The Trieste tunes of this model to LHC data are presented. The closepacking model is in qualitative agreement with many of the salient particle ratios, although the $Ξ_c/D$ ratio and the shape of $p_\perp$ spectra remain challenging to account for. Overall, the closepacking model with the Trieste tunes provides a competitive description of enhanced strangeness production in pp collisions, improving upon existing PYTHIA models while avoiding excessive proton yields.

Using the path-integral formalism, we show that photons possess a nontrivial quantum metric in momentum space. We derive the semiclassical action and equations of motion by taking into account the quantum metric. In media with a spatially varying refractive index $n(\mathbf{x})$, the quantum metric induces a shift in the trajectory of light at second order in derivatives of $n$, which may be regarded as a nonlinear Hall effect of light. The quantum metric also gives rise to corrections to gravitational lensing in curved spacetime at the nonlinear order in wavelength. This gravitational nonlinear Hall effect results from the interplay between the geometry of position space and that of momentum space.

The polarization state of light plays a central role in strong-field light--matter interactions and is widely used to probe electronic structure in solids via high-order harmonic generation (HHG). In particular, helicity-resolved HHG has been interpreted as a fingerprint of crystal symmetry and topology. Here, we demonstrate deterministic and continuous control of harmonic helicity in solids using polarization-crafted beams, formed by two orthogonally polarized pulses with a controlled time delay. By tuning this delay, the polarization state of individual harmonics can be driven from linear to circular, independent of the material under investigation. We show that this behavior is robust across systems with distinct symmetry and topology, and originates from the sub-cycle modulation of the light--matter interaction mediated by the dipole coupling. Furthermore, the orthogonal configuration allows to break the dynamical symmetry of the light-matter interaction which is manifested in the generation of otherwise forbidden harmonics under standard selection rules.. These results establish harmonic helicity as a field-controlled observable rather than a direct material fingerprint.

Isolated quantum systems typically approach thermal equilibrium as described by the Eigenstate Thermalization Hypothesis (ETH). Going beyond this involves either higher order correlators (full thermalization) or the formation of state designs, i.e., the approach of moments of state ensembles after a projective measurement towards thermal equilibrium (deep thermalization). We compare these two extensions of ETH using extensive numerical studies within a paradigmatic model for chaotic many-body quantum dynamics. For this we find exponential relaxation for both extensions: For deep thermalization all moments relax with the same rate, which approximately equals the relaxation rate of the autocorrelation function captured by ETH. In contrast, higher order correlation functions in full thermalization approach equilibrium faster. This means that at higher orders full thermalization is faster than deep thermalization.

We respond to the recent suggestion by A. Gal [arXiv:2604.18259] that the sharp pion-momentum peak at $p_{π^-} \approx 113.8$~MeV/$c$ observed in our $^7\mathrm{Li}(e,e^\prime K^+)$ electroproduction experiment at MAMI [Phys. Rev. Lett. 136, 152301 (2026)] originates from $^7_Λ\mathrm{He}$ weak decay rather than from $^3_Λ\mathrm{H} \to π^- + {}^3\mathrm{He}$ as we reported. We present quantitative arguments against this interpretation and conclude that the $^3_Λ\mathrm{H}$ assignment remains the most well-supported interpretation of the data.

We have developed a novel BEDT-TTF-based organic conductor, $κ$-(BEDT-TTF)$_2 m$-HOOCC$_6$H$_4$SO$_3$ ($κ$-$m$-SBA), and propose it as a candidate for altermagnet. Tight-binding band calculations of $κ$-$m$-SBA provide a $t'/t$ of 1.01 at 100 K, indicating that the spin structure is closely aligned to an equilateral triangle ($t'/t= 1$). While most $κ$-type BEDT-TTF-based salts become spin liquids due to the spin frustration caused by the triangular lattice, $κ$-$m$-SBA surprisingly shows a weak ferromagnetic transition at $T_\mathrm{N} = 14$ K due to a canted antiferromagnetic (AFM) spin structure. Until recently, $κ$-(BEDT-TTF)$_2$Cu[N(CN)$_2$]Cl ($κ$-Cl) was the only $κ$-type organic conductor known to exhibit this order, and it is also recognized as the first candidate for altermagnetism in organic conductors. This was theoretically predicted by Naka et al. in 2019, who demonstrated that $κ$-type organic conductors can be candidates for altermagnetism if they display such order. Consequently, $κ$-$m$-SBA can be considered the second candidate for altermagnetism in organic conductors. Furthermore, numerical calculations demonstrate a characteristic of altermagnets in $κ$-$m$-SBA, namely spin splitting of energy bands.

We investigate parameterized algorithms for computing the average-tree phylogenetic diversity (APD) in rooted phylogenetic networks, studying the problem under different structural parameters that capture the deviation of a network from a tree. Our primary parameter is the scanwidth, a measure of the tree-likeness of a given directed acyclic graph. We show that a subset of taxa with maximum APD can be found in polynomial time in phylogenetic networks of scanwidth at most 2, but becomes NP-hard in networks of scanwidth 3. Further, we design an algorithm that computes the APD of a given set of taxa in O(2^sw n) time, where sw denotes the scanwidth and n the number of taxa in the input network. Finally, we give a linear-time algorithm for computing the APD of a given set of taxa if the network induced by these taxa is reticulation-visible. We generalize this algorithm to still run in polynomial time if each biconnected component of the induced network has only constantly many invisible reticulations.

In this paper, we study the three-dimensional inviscid incompressible resistive Hall-MHD system in the axisymmetric setting with nontrivial swirl velocity and purely azimuthal magnetic. Assuming only that the swirl component of the initial velocity is sufficiently small, we prove that the lifespan of the strong solution can be sufficiently large. An explicit lifespan lower bound in terms of the size of the initial swirl is given. Moreover, we also study the behavior of the lifespan as the resistivity tends to zero.

We propose an OAM sorter based on a novel optical element that we refer to as a wavefront twister. It is a generalization of the conventional wavefront rotators such as the Dove prism. However, unlike a Dove prism, which simply rotates a wavefront, the rotation generated by a wavefront twister varies linearly with radial position, resulting in the twisting of the wavefront. We demonstrate that the wavefront twister, followed by a lens, maps each OAM mode to an annulus of distinct radius at the back focal plane of the lens with negligible inter-modal overlap and preserves the circular symmetry. Thus, the proposed wavefront twister offers a scalable scheme for high-dimensional OAM mode sorting, with important consequences for the practical realization of OAM-based applications.

Crowds have long held a paradoxical place in the human imagination, feared for their destructive potential yet essential for collective expression. This tension was tragically manifested in the 1919 Jallianwala Bagh massacre, when British colonial troops opened fire on a peaceful gathering in Amritsar, India. Although officially 379 deaths were recorded, eyewitnesses and historians have long challenged this figure. With this study, we critically revisit the events through the lens of the specific role of the crowd as a phenomenon, both regarding the physical and the socio-psychological dynamics. We show that even under conservative physical assumptions - moderate shooting cadence, crowd-shielding, and constrained escape routes - our agent-based simulations consistently yield fatality estimates well above the official death count. On the socio-psychological front, we explore how early 20th-century discourses, influenced by Le Bon's theory of crowd psychology, constructed the crowd as an inherently irrational and threatening entity, thus providing a rationale for the application of excessive force. Our findings show that acknowledging the socio-cultural construction of crowds as a relevant factor in how state power engage with and respond to collective gatherings brings to light contemporary parallels and the risks posed by their rhetorical framing. Furthermore, this study highlights the importance of interdisciplinary modelling for both historical accountability and current crowd safety, particularly in an era of growing political unrest, surveillance, and militarised crowd policing.

The superconducting diode effect (SDE) allows polarity-dependent critical currents when time-reversal and current-inverting spatial symmetries are broken. Superconducting diodes show promise for applications, but inversion asymmetry is usually encoded in sample geometry or non-centrosymmetric crystals, rendering them static circuit elements. Here we demonstrate a programmable superconducting diode whose functionality is encoded in correlated electronic domains. We use the nematic superconductor FeSe as a platform and report a large intrinsic SDE with efficiencies up to $η\sim 75\%$ due to vortices interacting with nematic twin boundaries. The domain wall configuration thus encodes the SDE of the device. Through intense microsecond current pulses to quench the nematic order at rates exceeding $10^7$ K/s, we modify the domain pattern and control the polarity and strength of the SDE. These results establish a new paradigm in which superconducting circuit elements can be programmed through patterns imprinted into correlated electronic states.

Context. Growing evidence is revealing the crucial role of binarity in massive star evolution. This affects evolution models and demands a refinement of the available observational constraints. Aims. To investigate the possible evolutionary origins of a sample of 117 Galactic O-type stars with luminosity classes V to III and projected rotational velocities below 150 ${\rm km}\,{\rm s}^{-1}$. Methods: We extend previous quantitative spectroscopic analyses performed within the framework of the IACOB project and obtain N abundance estimates. We investigate correlations between these abundances and other stellar parameters. As a reference, we use predictions from single-star evolution models computed using different physical prescriptions. Results. We identify differences in the N abundance distributions corresponding to three He abundance regimes (He-low, He-normal, and He-rich). For the He-normal group, the N abundance distribution peaks slightly above the expected birth value and extends up to $ε_{\rm N}$=8.4 dex. For these stars, we find overall agreement with single-star evolutionary models that include efficient internal mixing and assume moderate-to-low initial rotation. In contrast, the He-rich group exhibits a bimodal N abundance distribution, with one peak at $\sim$8.1 dex and a second more enriched peak around $\sim$8.5 dex; none of these stars are consistent with predictions from single-star evolutionary models. We argue that He-rich stars are most plausibly explained as binary products. Furthermore, despite N abundance in He-normal stars with LC IV and V are reproduced by single stellar evolutionary models with efficient mixing, the same models predict a higher N abundance than observed for stars with LC III. This indicates that rotational mixing alone is unable to explain the observed distribution of N abundances among stars with normal He abundances.

We prove that over a generic Poncelet triangle family, the locus of the circumcenter of an inversive triangle is a conic. Additionally, we prove an earlier conjecture: over generic Poncelet triangles, two unique points exist which maintain constant power with respect to the circumcircle and Euler's circle of the family, respectively.

From an identity connecting a combinatorial sum and Legendre polynomials, we derive closed forms for a number of combinatorial sums. Some of them are obtained via results about the integrals of functions associated with Legendre polynomials.

We study the problem of distributed control of large-scale robotic swarms which can be modeled as continuum densities evolving under the continuity equation. We propose a formalization of distributed controllers as (generally nonlinear) differential operators, in which control inputs depend only on local information about the state and environment. This perspective yields a fully local, PDE-based framework for analysis and design. We apply this framework to the problem of stabilizing a swarm density around an arbitrary target density, and investigate fundamental limitations of low-order distributed controllers in achieving this goal. In particular, we show that controllers which act in a purely pointwise manner are incompatible with natural system symmetries and strong forms of stability, and must rely on mixing-type behavior to achieve stabilization. In contrast, we present a simple first-order control law which achieves stabilization and enjoys substantially stronger properties.

The discrete Boltzmann factor $B_E(β_n)=(1-bE)^n$, introduced by Chung, Hassanabadi, and Boumali, provides a lattice regularization of the canonical weight $e^{-βE}$ and imposes the compact-support condition $E<1/b$. In the present analysis we systematically separate results that follow directly from this bounded thermal weight from those that require additional phenomenological input. First, we study the discrete Bose--Einstein occupation factor relevant for Hawking radiation, derive the leading suppression of black-hole luminosity, and show that the thermal Hawking channel shuts off as the cutoff scale is approached. Second, we formulate a discrete work functional built from ratios of thermal weights and establish an exact Jarzynski-type identity for deterministic measure-preserving protocols; in contrast, the corresponding Crooks relation does not collapse to a function of work alone, and first-order approximations retain an explicit initial-energy dependence that cannot be reduced to a simple $W$-dependent correction without additional assumptions. Third, and purely as an ancillary kinematic extension rather than a derivation from the statistical framework itself, we examine a bounded modified-dispersion ansatz and estimate the associated time-of-flight constraints. Throughout, we include illustrative figures, clarify the non-universal status of the entropy correction, and emphasize that direct laboratory signatures are negligible whenever $b$ is universal and Planck suppressed. Finally, the standard continuum expressions are recovered smoothly in the limit $b\to 0$.

The reliability of machine learning in multiscale physical systems depends on how physical structure is embedded into the learning process. We investigate this in the context of turbulent multiphase flows, focusing on the prediction of subgrid interfacial area density, a key quantity governing interphase transport that remains unresolved in large-eddy simulations. In this work, we develop and evaluate two machine learning subgrid closure models to predict the three-dimensional subgrid interfacial area density: a purely data-driven 3D encoder-decoder network, and a physics-constrained variant regularized by a fractal geometric prior. Across a range of Weber numbers, the physics-based model improves predictive accuracy, reduces error variance, and suppresses nonphysical artifacts relative to purely data-driven approaches. We also show that these gains are regime-dependent: the embedded inductive bias enhances generalization in corrugation-dominated regimes where its underlying assumptions hold, but becomes ineffective in fragmentation-dominated regimes characterized by topology change and droplet breakup. These results reveal a broader principle for scientific machine learning: the utility of physics-informed models depends not only on the presence of inductive bias, but on its alignment with the governing physical regime. This suggests a path toward regime-aware learning frameworks for modeling of complex multiscale systems.