PhysInOne: Physics-Informed Models & Datasets

Updated 4 July 2026

PhysInOne is a term for distinct physics-centered artifacts: a cardiovascular PINN framework and a large-scale visual physics dataset.
In cardiovascular modeling, it integrates a closed-loop ODE system with neural networks to rapidly estimate parameters from LV pressure-volume data with <5% error.
In visual physics, it offers 2 million videos across 153,810 scenes to benchmark physics-aware video generation, future prediction, and physical property estimation.

PhysInOne is a name used in the arXiv literature for more than one physics-centered machine-learning artifact. In "Rapid Estimation of Left Ventricular Contractility with a Physics-Informed Neural Network Inverse Modeling Approach" it denotes a physics-informed neural-network framework for cardiovascular parameter estimation built around a closed-loop lumped-parameter model of the circulation, with forward and inverse modes focused on the left ventricle (LV) and left atrium (LA) (Naghavi et al., 2024). In "PhysInOne: Visual Physics Learning and Reasoning in One Suite" it denotes a large synthetic dataset for visual physics learning and reasoning, containing 2 million videos across 153,810 dynamic 3D scenes and covering 71 basic physical phenomena (Zhou et al., 10 Apr 2026). This suggests that the term is best interpreted from disciplinary context: in one usage it names a PINN-based cardiovascular inverse model, and in another it names a benchmark-and-training suite for physics-grounded visual learning.

1. Multiple referents of the name

In the supplied literature, the two primary uses of the name are distinct in ontology, scale, and intended task. One is a physiology-constrained inference system for hemodynamic parameter estimation; the other is a synthetic corpus for training and evaluating visual world models.

Use of the name	Domain	Core role
PhysInOne	Cardiovascular modeling	Physics-informed neural-network framework for cardiovascular parameter estimation
PhysInOne	Visual physics	Large-scale synthetic dataset for visual physics learning and reasoning

The ambiguity is substantive rather than terminological. In the cardiovascular work, PhysInOne is explicitly the authors’ framework for embedding a cardiovascular ODE model into a neural network so that network outputs satisfy governing physiology while also matching data (Naghavi et al., 2024). In the visual-physics work, PhysInOne is a dataset with scene generation, rendering, annotations, and benchmarks for downstream applications (Zhou et al., 10 Apr 2026).

The supplied literature also contains several related physics-informed systems presented under other names: ONION for line-integral diagnostics across fusion devices, PI-Latent-NO for high-dimensional parametric PDE systems, BP-DeepONet for cuffless arterial blood-pressure estimation, a Physics-Informed Neural ODE framework for cardiac $T_1$ mapping, and PCNDE for one-dimensional blood-flow surrogates (Wang et al., 2024, Karumuri et al., 14 Jan 2025, Li et al., 2024, Capitão et al., 1 Jul 2025, Csala et al., 2024). A plausible implication is that “PhysInOne” sits within a broader research tendency toward explicit integration of governing physics into representation learning, inverse modeling, and operator learning.

2. PhysInOne as a cardiovascular PINN framework

In the 2024 cardiovascular usage, PhysInOne is a PINN built around a closed-loop blood circulation system embedding a left ventricle, trained to satisfy a system of ODEs associated with a lumped-parameter description of the circulatory system (Naghavi et al., 2024). The underlying model contains compartments for the left ventricle, left atrium, aortic valve, systemic arteries, and systemic veins, connected by resistive and compliant elements. The stated closed-loop parameters include valve and vascular resistances $R_{av}$ , $R_{ao}$ , $R_{art}$ , $R_{vc}$ , $R_{mv}$ ; vascular compliances $C_{ao}$ , $C_{art}$ , $C_{vc}$ ; total blood volume and reference volumes $V_{total}$ , $R_{av}$ 0, $R_{av}$ 1, $R_{av}$ 2; and cardiac period $R_{av}$ 3.

The LV constitutive law uses an exponential end-diastolic pressure–volume relation with $R_{av}$ 4, $R_{av}$ 5, and $R_{av}$ 6. LV activation uses $R_{av}$ 7, $R_{av}$ 8, $R_{av}$ 9, and $R_{ao}$ 0. The LA has analogous parameters: $R_{ao}$ 1, $R_{ao}$ 2, $R_{ao}$ 3, $R_{ao}$ 4, $R_{ao}$ 5, $R_{ao}$ 6, and $R_{ao}$ 7.

The PINN encodes conservation of volume in each compartment, flow through each valve or vascular segment driven by pressure differences and resistance, time-varying chamber pressure from elastance/activation laws, and periodic boundary conditions over one beat. In forward mode, the network takes time as input and predicts state trajectories, while the loss enforces data matching at sampled time points, ODE residual minimization at collocation points, and consistency with the closed-loop physiology. The reported forward-model result is that predictions have a maximum error of less than 5% when compared to numerical ODE solutions.

The inverse mode is the paper’s central application. PhysInOne takes measured or synthetic single-beat LV pressure and volume waveforms and optimizes the unknown parameter set so that the PINN-generated trajectories satisfy both the observed data and the governing ODEs. The inverse modeling approach estimates model parameters in $R_{ao}$ 8 mins from single-beat LV pressure and volume waveforms. Using synthetic LV pressure and volume waveforms generated by the PINN with different parameter values, the approach can recover the corresponding ground-truth values, and the paper states that this suggests that the model parameters are unique.

The experimental application uses 11 swine models, including data acquired before and after administration of dobutamine in 3 animals. The estimated LV end-systolic elastance $R_{ao}$ 9 is reported as about 58% to 284% higher for the data associated with dobutamine compared to those without, which the paper interprets as evidence that the framework can estimate LV contractility from single-beat measurements. A common misconception would be to reduce this PhysInOne to a generic PINN; in the paper it is specifically a forward–inverse cardiac modeling system tied to a closed-loop lumped-parameter circulation and to contractility estimation from LV pressure–volume data.

3. PhysInOne as a visual-physics dataset

In the 2026 usage, PhysInOne is a large synthetic dataset built to teach and evaluate AI systems on visual physics, understood as predicting and reasoning about how objects move and interact under physical laws (Zhou et al., 10 Apr 2026). The dataset is constructed in response to the claim that most existing physics datasets are small, narrow in scope, and often contain simplified scenes with only one kind of motion or a single object.

Its scale is unusually large by the standards stated in the paper: 2 million videos, 153,810 dynamic 3D scenes, 71 basic physical phenomena, and 3,284 multiphysics activities derived from combining those phenomena. Scenes are built from 2,231 3D objects, 623 materials, and 528 3D backgrounds. Each scene is rendered from 13 videos: 12 fixed cameras and 1 moving monocular camera. Rendering is at $R_{art}$ 0 resolution, usually at 30 FPS, with an average duration of 5.2 seconds; roughly 75% of the videos are longer than 5 seconds.

The phenomena span four domains: mechanics, optics, fluid dynamics, and magnetism. The paper explicitly omits thermodynamics and acoustics. Representative phenomena listed in the paper include object collisions, rolling on slopes, wind acting on objects, magnetic attraction and repulsion, rotating systems, springs, ropes, seesaws, pendulums, catapults, buoyancy, reflection and laser-beam interactions, fluid transfer, surface tension, Newtonian and non-Newtonian fluids, and granular behavior.

A defining property of the dataset is scene difficulty. The paper emphasizes multiobject interactions, multiphysics activities, complex objects such as destructible, deformable, granular, and liquid objects, and complex backgrounds including rooms, factories, pools, kitchens, and mountains. Average object count rises with activity complexity: 3.9 objects per scene for single-physics activities, 6.3 for double-physics activities, and 7.8 for triple-physics activities.

Supervision is correspondingly rich. Geometry annotations include depth maps in meters and 3D meshes for all assets. Semantic annotations include per-frame segmentation masks, with each foreground object assigned a unique ID and background labeled as 0. Motion annotations include 3D trajectories and object pose over time, including position and rotation. Physical-property annotations include dynamic friction coefficient, static friction coefficient, density, restitution coefficient, and, in the property-estimation benchmark, material-specific parameters such as Young’s modulus $R_{art}$ 1, Poisson’s ratio $R_{art}$ 2, yield stress $R_{art}$ 3, viscosity $R_{art}$ 4, bulk modulus $R_{art}$ 5, plasticity viscosity $R_{art}$ 6, and friction angle $R_{art}$ 7. Each scene also has a manually curated paragraph of text, polished with a LLM, with average length about 64 English words per scene.

Dataset construction uses Unreal Engine 5 / Chaos Physics for most rigid-body and interaction dynamics, Taichi / MPM for deformable and granular simulations, and SPH via Doriflow for fluid simulation. Controlled parameter ranges include friction coefficient $R_{art}$ 8, restitution $R_{art}$ 9, density $R_{vc}$ 0, fluid viscosity $R_{vc}$ 1, plasticity viscosity $R_{vc}$ 2, and friction angle $R_{vc}$ 3. The data split is 8:1:1, and 3D assets are kept exclusive to one split.

4. Benchmark suite and empirical findings in visual physics

PhysInOne is not only a repository of videos; it is also presented as a benchmark suite for four application classes: physics-aware video generation, long-/short-term future frame prediction, physical property estimation, and motion transfer (Zhou et al., 10 Apr 2026).

For physics-aware video generation, the paper fine-tunes SVD-XT, CogVideoX-1.5-5B, and Wan2.2-5B using LoRA, SFT, and FLT, on a subset of 83,650 text-video pairs. To evaluate motion quality, it introduces Physical Motion Fidelity (PMF), defined from the frequency-domain energy distribution of generated and reference videos: $R_{vc}$ 4 with

$R_{vc}$ 5

The paper states that fine-tuning on PhysInOne improves physical plausibility substantially, that human judgments correlate well with PMF, and that models tend to do better in fluid and magnetism than in mechanics and optics.

For future-frame prediction, long-term prediction uses the first half of a video to predict the second half, roughly $R_{vc}$ 6 seconds ahead and about 78 future frames. Tested methods include TiNeuVox, DefGS, FreeGave, TRACE, ExtDM, and MAGI-1, evaluated with PMF, PSNR, SSIM, and LPIPS. Short-term prediction asks the model to predict the next 10 frames continuously at each time step. The reported finding is that current methods perform reasonably on seen viewpoints but drop on novel viewpoints, indicating difficulty in modeling 3D physical evolution and long-horizon dynamics.

For physical property estimation, the paper benchmarks PAC-NeRF and GIC on elastic solids, plasticine, Newtonian fluids, non-Newtonian fluids, and granular substances. Both methods can infer physically plausible properties, but they often fail on harder scenes with complex shapes, complex backgrounds, and complicated multiobject interactions. The paper also evaluates resimulation using the estimated parameters and reports that GIC generally performs better than PAC-NeRF, while both remain imperfect.

For motion transfer, the benchmark uses 273 paired scenes and evaluates MotionPro and GoWithTheFlow. The reported result is that these methods can generate visually realistic outputs but fail to transfer complex physical motion faithfully. The central message across the four tasks is consistent: foundation models improve when exposed to PhysInOne, but they still struggle with true physical reasoning, especially under complex dynamics, novel views, long horizons, and inverse-physics settings.

5. PhysInOne as a benchmark for 3D-aware world models

The dataset has also become an evaluation target for models that attempt to learn physics from video without explicit simulator states. "Neural Voxel Dynamics: Learning Implicit 3D Physics via Volumetric Feature Advection" evaluates a self-supervised 3D latent model on PhysInOne and describes the benchmark as a more complex multi-phenomena physics benchmark with ground-truth depth, 12 stationary cameras, ~150-frame trajectories, multiple simultaneous physical phenomena, and interactions including rigid-body collisions, liquid dynamics, and wind-driven motion (Wang et al., 24 Jun 2026).

In that evaluation, future latent prediction is measured in V-JEPA feature-space L2 loss, with both 2D latent L2 and 3D latent L2 reported. The strongest single-camera GT-depth result on PhysInOne is Ours-GT: $R_{vc}$ 7, compared with CogVideoX: $R_{vc}$ 8, PhysGen: $R_{vc}$ 9, PhysGaussian: $R_{mv}$ 0, PhysCtrl: $R_{mv}$ 1, and a 2D baseline: $R_{mv}$ 2. Under single-camera estimated depth, Ours-estimate is reported as $R_{mv}$ 3.

Multi-camera results reinforce the same point. With GT depth, Ours-GT gives $R_{mv}$ 4, and with estimated depth, Ours-estimate gives $R_{mv}$ 5. The paper interprets these results as evidence that accurate geometric lifting into a 3D latent volume is a major factor in success on PhysInOne, while estimated-depth quality remains a bottleneck. Supporting ablations show that MoGe yields IoU 87.13, 2D feat 1.78, 3D feat 0.61, outperforming VGGT and VidDepA + Mast3r, and that adding occupancy and observed/unseen channels improves performance from feature only: 4.97 / 2.42 to feat + occ + obs: 1.78 / 0.61.

These results are significant because they place PhysInOne in a role beyond dataset curation: it functions as a stress test for whether learned models preserve 3D structure, object permanence, and heterogeneous dynamics under multi-view and long-horizon conditions. A plausible implication is that the dataset has become a discriminative benchmark precisely because its complexity exceeds what purely 2D latent predictors can encode reliably.

6. Conceptual relations to other physics-informed learning systems

Although only two artifacts in the supplied literature are formally named PhysInOne, several other systems exhibit the same architectural principle: physical laws are not treated as post hoc plausibility checks but are inserted into representation, fusion, or optimization itself.

In fusion diagnostics, ONION injects the geometry of the diagnostic system and the line-integration law into both network structure and training objective. Its distinctive mechanism is element-wise multiplication between learned features and encoded physical information, rather than addition or concatenation, and its physics-informed loss reprojects predicted emissivity fields through the known contribution matrix $R_{mv}$ 6 (Wang et al., 2024). In operator learning, PI-Latent-NO couples a Latent-DeepONet and a Reconstruction-DeepONet, enforcing PDE residuals, boundary conditions, and initial conditions through automatic differentiation while exploiting time–space separability for scalability (Karumuri et al., 14 Jan 2025). In cuffless hemodynamics, BP-DeepONet predicts a continuous pressure and flow field $R_{mv}$ 7, $R_{mv}$ 8 while satisfying the 1-D Navier–Stokes equations, a time-periodic condition, and a three-element Windkessel boundary condition (Li et al., 2024). In cardiac MRI, a continuous-time LSTM-ODE is combined with physics-informed losses derived from the inversion-recovery signal model to estimate voxel-wise $R_{mv}$ 9 from 3–5 baseline images instead of the conventional 11-image MOLLI acquisition (Capitão et al., 1 Jul 2025). In one-dimensional blood-flow surrogates, PCNDE rewrites the momentum equation as a spatial neural PDE, $C_{ao}$ 0, using a space-time swap to exploit temporal periodicity and improve coupling stability (Csala et al., 2024).

A related but methodologically distinct endpoint is physics-encoded spatio-temporal regression, where the model class itself is defined on the PDE solution manifold rather than regularized toward it. There the spacetime field is expanded as

$C_{ao}$ 1

so time evolution is hard-wired through eigen-decay factors $C_{ao}$ 2 (Li et al., 2024). This differs from PINN or DeepONet practice, but it shares the same commitment to constraining inference by governing dynamics.

Taken together, these systems clarify what is specific and what is general about PhysInOne. What is specific is the object being named: a cardiovascular PINN in one paper, a visual-physics dataset in another. What is general is the research program they exemplify: the use of closed-form physics, ODE/PDE residuals, constitutive relations, forward operators, or simulator-grounded annotations as first-class elements of machine-learning system design.

7. Interpretation, scope, and common misconceptions

The main interpretive difficulty surrounding PhysInOne is nominal rather than technical: the same label refers to heterogeneous artifacts across subfields. In cardiovascular modeling it is not a benchmark dataset, and in visual physics it is not a PINN-based inverse model. The two usages share a physics-centered design philosophy, but they operate at very different levels of abstraction.

Another possible misconception is to equate the cardiovascular PhysInOne with unconstrained black-box regression. The 2024 work explicitly encodes a closed-loop ODE system, uses single-beat LV pressure and volume waveforms as observations, and evaluates uniqueness through Sobol sensitivity analysis with first-order, second-order, and total-order indices (Naghavi et al., 2024). Likewise, it would be inaccurate to treat the 2026 PhysInOne merely as a larger video dataset. Its defining features include multiobject and multiphysics complexity, five classes of annotations, multiple rendering views, explicit benchmarks, and the introduction of PMF as a motion-quality measure (Zhou et al., 10 Apr 2026).

The broader literature indicates that these two usages participate in adjacent but distinct agendas. One agenda emphasizes rapid inverse modeling and parameter estimation in physiology; the other emphasizes physics-grounded data curation for world models, generation, prediction, and inverse physics. This suggests that “PhysInOne” is best read as a context-dependent label attached to the integration of physics with learning, rather than as the name of a single unified method, dataset, or software stack.