Pith. sign in

REVIEW 4 major objections 4 minor 36 references

Physically Constrained 3D Diffusion for Inverse Design of Fiber-reinforced Polymer Composite Materials

T0 review · 4 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A 3D diffusion model can now inverse-design fiber-reinforced composites to match a target nonlinear stress-strain curve, with physical constraints enforced during generation.

desk verdict A useful, largely reproducible diffusion pipeline for composite inverse design, but the 'entire stress-strain curve' claim rests on only three strain points and should be scoped down. read the letter →

arxiv 2412.01321 v1 pith:D7Q652BL submitted 2024-12-02 cond-mat.soft cond-mat.mtrl-sci

Pei Xu , Yunpeng Wu , Srikanth Pilla , Gang Li , Feng Luo This is my paper · ORCID
classification cond-mat.softcond-mat.mtrl-sci
keywords fiber-reinforcedpolymercompositesinversedesigndiffusionmodelstress-straincurvecollision-freegenerationphysicalconstraintsgenerativeFEAsimulation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper reports a generative design method, PC3D_Diffusion, that tackles the inverse problem of fiber-reinforced polymer composites: given a desired nonlinear stress-strain curve, it produces a three-dimensional fiber arrangement whose simulated mechanical response closely matches that curve. The authors generate and simulate 1.35 million composite samples to train a 3D diffusion model over fiber positions and orientations. A vanilla diffusion model produces visually plausible but mostly intersecting fibers, so they add a differentiable collision-loss guidance during the reverse generative process. The result is a generator that produces collision-free designs with stress-strain curves close to the target, on average within one percent, and it also works for configurations and curves outside the training range. If it holds up, this would be the first inverse design of fiber composites targeting the whole stress-strain curve rather than a single stiffness or strength value.

What carries the argument

The key machinery is a 3D spatial diffusion model over a variable set of fibers, where each fiber is represented by a position and an orientation treated as a point in a cylinder-symmetric rotation space. Position noise is added in variance-preserving form, orientation noise in a variance-exploding rotation form, and a transformer-decoder network with a graph-attention encoder estimates both noises conditioned on the target curve coefficients and fiber geometry. Physical feasibility is enforced by a differentiable collision constraint loss that measures pair-wise fiber distances; during the reverse denoising process this loss is minimized in a learning-free gradient loop, so the final sample is collision-free without retraining.

What would settle it

Pick a target curve, generate a fiber distribution with PC3D_Diffusion, then run finite-element analysis at many strain increments (for example every 2% from 0 to 30%) and compare the full simulated stress-strain curve against the target cubic; if the full-curve deviation substantially exceeds the error reported on the three fitted points, the cubic three-point fit is the failing simplification.

Watch

Extended reading notes

Core claim

The paper's central claim is that PC3D_Diffusion can perform inverse design of fiber-reinforced composites against the full nonlinear stress-strain curve: the user provides the coefficients of a target cubic stress-strain curve and a fiber configuration (diameter, length), and the model returns fiber position/orientation distributions whose FEA-simulated curves match the target. On held-out configurations, the best-of-10 generated designs have mean absolute errors below 0.5% in all curve coefficients and relative curve-area errors below 0.2%; average-of-10 errors remain below 1% and 0.45% respectively. The same pipeline also generates composites for fiber configurations and target curves that lie outside the training envelope. The authors further show that the loss-guided constraint layer guarantees collision-free fiber distributions while barely changing the overall generated microstructure.

Load-bearing premise

The entire nonlinear stress-strain response is represented by a cubic polynomial fitted to only three simulated strain points (10%, 20%, 30%), and the deterministic finite-element analysis used to generate all training labels is treated as the true mechanical response of the composite.

Editorial extensions

If this is right

  • Designers can target an entire mechanical response curve rather than a single property, which matters for energy-absorbing and soft-robotic applications that rely on nonlinear deformation behavior.
  • Multiple diverse but equivalent fiber distributions can be generated for the same target curve, giving engineers options to trade off manufacturability and cost.
  • The same diffusion-plus-guidance framework extends to other 3D microstructures and other physics-based objectives, such as particulate or laminar composites and damage or fatigue properties.
  • The reported generalization to out-of-range curves suggests the method can propose composite designs with stress-strain behavior not present in the training data.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the model is trained entirely on deterministic FEA labels, its accuracy outside simulation depends on how faithfully that FEA model reproduces real composites; physical validation on manufactured samples would be the decisive test.
  • The paper implicitly assumes that a cubic polynomial fitted to three strain points captures the full curve; if the true response has sharper nonlinearity between those points, the reported accuracy may not reflect complete curve matching.
  • A natural next step is to add a feasibility check at inference time that rejects target curves with no realizable microstructure, which would make the inverse design tool more reliable for engineering use.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper presents PC3D_Diffusion, a conditional 3D spatial diffusion model for the inverse design of fiber-reinforced polymer composites. The model is trained on 1.35 million simulated FRPC microstructures whose stress-strain curves are obtained by FEA at 10%, 20%, and 30% strain and represented as zero-intercept cubic polynomials. During generation, a learning-free loss-guided correction is applied in the reverse diffusion process to enforce collision-free fiber arrangements. Quantitative evaluation on held-out samples reports best-of-10 relative MAEs below 0.5% and average MAEs below 1% for the cubic coefficients, with relative curve-area errors E_a below 0.2% (best-of-10) and below 0.45% (average), plus a claim of exact collision-free generation. The paper also demonstrates qualitative generalization to configurations and target curves outside the training envelope, and it discusses limitations related to manufacturability, feasibility, and the deterministic FEA surrogate.

Significance. If the central claims hold, this is a substantial methodological advance: it is the first reported inverse-design framework for FRPCs that targets the full stress-strain response rather than a single stiffness metric, and the proposed collision-avoidance guidance is a simple, learnable-free mechanism that could transfer to other 3D generative design tasks. The paper's strengths include the release of a large dataset and code, the use of both best-of-10 and average-error reporting, ablations of the graph-attention representation, and an explicit statement of limitations. The main weakness is that the 'entire nonlinear stress-strain curve' claim is supported only at three strain points, and the 'guarantee' of collision-free generation is stated more strongly than the approximate distance loss justifies. These issues are fixable with additional validation and more careful wording.

major comments (4)
  1. [Methods (Database Generation), Eq. (1), Eq. (14)] The paper's central claim that PC3D_Diffusion performs inverse design of the 'entire nonlinear stress-strain curve' is not supported by the reported metrics. The FEA pipeline computes stress at only three strain levels (10%, 20%, and 30%), and the curve is represented by the zero-intercept cubic σ(ε) = a1 ε + a2 ε² + a3 ε³. Because this cubic has exactly three free parameters, matching (a1, a2, a3) is mathematically equivalent to matching stress at the three fitted strain points; the relative MAE and E_a in Eq. (14), which integrates the two fitted cubics over [0.1, 0.3], cannot detect non-cubic response at intermediate strains such as 15% or 25%. To support the 'entire curve' claim, the authors should verify that the cubic interpolation error is small by running FEA at additional strain increments on a subset of generated samples, and either report those checks or rephrase the claim as inverse design at three sampled strain points.
  2. [Abstract and Section Physical Constrained Generation, Eq. (1)] The word 'guarantee' overstates what the differentiable collision loss provides. The distance function and the length-shrinking function in Eq. (1) are approximations (as described in Supplementary Notes), so reaching L_cons = 0 implies non-penetration for the shrunk, smooth representation, not necessarily for the original finite-element cylinders. Moreover, the main text says collision-free results are achieved 'typically within 10 iterations,' which is an empirical observation rather than a guarantee. Either provide a certificate based on exact geometry or replace 'guarantee' with a more precise statement of the conditions under which L_cons = 0 is necessary and sufficient for non-penetration of the original fibers.
  3. [Section Quantitative Evaluation and Discussion] All verification uses the same deterministic FEA pipeline that generated the training labels, so the reported MAE and E_a values quantify self-consistency of the surrogate model rather than accuracy relative to physical composite behavior. The Discussion explicitly acknowledges this limitation, but the Abstract and Results currently present the accuracy as intrinsic to the method. The authors should qualify the headline numbers (e.g., 'within the FEA surrogate') or add a small validation set using an independent simulator or experimental measurements.
  4. [Section Generation using Out-of-the-Range Conditions, Fig. 4] The out-of-range generalization claim is supported only by visual comparison of curves; no quantitative errors (MAE or E_a) are reported for the four in-range-configuration cases or the two below-envelope cases. Since generalization beyond the training range is one of the paper's main claims, at least the below-envelope cases should be accompanied by the same error metrics used in Table 1, or by FEA-verified stress values at all three strain points.
minor comments (4)
  1. [Results and Methods (Database Generation)] The total dataset size is stated as '1.35 million' in the Abstract and Methods but as '1.3 million' in the Results; 52 × 26,000 = 1,352,000, so the latter is incorrect and should be harmonized.
  2. [Equation (5) and surrounding text] The condition vector c in Eq. (5) is used but not explicitly defined in the main text; the authors should state that c contains the curve coefficients a1, a2, a3 and the fiber configuration variables d and l, while n enters through the number of input fiber tokens.
  3. [References and typesetting] Several equations and references are garbled in the present version (e.g., reference [32] begins with an unreadable string, and Eq. (12) contains stray glyphs); the camera-ready version should be typeset cleanly and all references should be checked.
  4. [Section Physical Constrained Generation] The phrase 'completely collision-free results typically with less than 10 iterations' would be clearer if accompanied by the empirical distribution of the number of refinement iterations needed, especially for the densest configurations (n=50) where the initial collision-free percentage is the lowest.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the learned inverse mapping is validated on held-out FEA samples, and the three-point cubic curve representation is a stated modeling choice rather than a self-referential derivation.

full rationale

The central derivation is a learned conditional diffusion model: the paper trains a 3D spatial diffusion model on 1.35 million FEA-generated fiber composites, conditions generation on the coefficients (a1, a2, a3) of a target stress-strain curve, and validates on held-out test samples from configurations not seen in training. This is supervised inverse mapping, not equation-level circularity. The paper's stress-strain curves are defined as zero-intercept cubics fitted to FEA stresses at 10%, 20%, and 30% strain (Methods, Database Generation), so the reported coefficient MAE (Eq. 13) and area error E_a (Eq. 14) only certify agreement on those three strain points and on the cubic interpolants between them. This narrows the 'entire stress-strain curve' claim, but it is not circular: the FEA stress values are independent outputs of a deterministic simulator, and the model can fail to match them. The collision-free guarantee is likewise an explicit constraint-enforcement procedure, minimizing L_cons by gradient descent until it reaches zero (Eqs. 1-2), rather than a fitted quantity renamed as a prediction. Self-citations are limited to background deep-reinforcement-learning material design work (e.g., Ref. 18) and do not carry the argument; there is no imported uniqueness theorem or ansatz forcing the model choice. The Discussion openly acknowledges that the training data come from deterministic FEA and that experimental validation and stochastic effects are out of scope. These are external-validity limitations, not derivation-level circularity.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim depends on three families of assumptions: the FEA simulator as ground truth, the cubic-coefficient representation of the whole curve, and convergence of the differentiable collision loss. No new particles, forces, dimensions, or conserved quantities are introduced.

free parameters (4)
  • Stress-strain cubic coefficients (a1, a2, a3) = Per-sample fitted values from FEA
    The target curve is summarized by three coefficients fitted to stress at 10%, 20%, and 30% strain; every generation condition and every error metric uses these fitted coefficients.
  • Diffusion time steps and variance schedule endpoints (T, beta range, sigma range) = T=500; schedule endpoints stated in Methods
    These hand-chosen hyperparameters control the forward and reverse diffusion processes and affect generation quality.
  • Multi-task loss weights lambda_p and lambda_r = Learned during training
    Equation (9) introduces learnable weights to balance position and orientation noise losses; they are trained, not physical constants.
  • Minimal gap p in collision loss = Not numerically specified in extracted text
    The collision-free definition and the strength of the physical constraint depend on this chosen gap value; smaller or larger p changes what counts as valid.
assumptions (3)
  • domain assumption Deterministic FEA using hyperelastic matrix model and linear fibers yields the true stress-strain response of the generated microstructures.
    All training labels and all final verification use this FEA pipeline; no experimental measurements are reported.
  • ad hoc to paper The full nonlinear stress-strain curve is adequately represented by a cubic polynomial fitted to three simulated stress points at 10%, 20%, and 30% strain.
    This cubic representation is both the target condition and the basis for the MAE and curve-area error metrics.
  • ad hoc to paper Fiber collisions are fully described by the differentiable distance loss with a minimal gap parameter and shrunken fiber length, and reaching zero of this loss guarantees collision-free generation.
    The 'guarantee' claim rests on convergence of gradient descent; the paper itself reports typical rather than universal convergence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Physically Constrained 3D Diffusion for Inverse Design of Fiber-reinforced Polymer Composite Materials." pith.science (2026). https://pith.science/paper/D7Q652BL

@misc{pith2026241201321,
  author       = {Pith},
  title        = {Pith review of: Physically Constrained 3D Diffusion for Inverse Design of Fiber-reinforced Polymer Composite Materials},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D7Q652BL}},
  note         = {Machine review of arXiv:2412.01321}
}
read the original abstract

Designing fiber-reinforced polymer composites (FRPCs) with a tailored nonlinear stress-strain response can enable innovative applications across various industries. Currently, no efforts have achieved the inverse design of FRPCs that target the entire stress-strain curve. Here, we develop PC3D_Diffusion, a 3D spatial diffusion model designed for the inverse design of FRPCs. We generate 1.35 million FRPCs and calculate their stress-strain curves for training. Although the vanilla PC3D_Diffusion can generate visually appealing results, less than 10% of FRPCs generated by the vanilla model are collision-free, in which fibers do not intersect with each other. We then propose a loss-guided, learning-free approach to apply physical constraints during generation. As a result, PC3D_Diffusion can generate high-quality designs with tailored mechanical behaviors while guaranteeing to satisfy the physical constraints. PC3D_Diffusion advances FRPC inverse design and may facilitate the inverse design of other 3D materials, offering potential applications in industries reliant on materials with custom mechanical properties.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 31 canonical work pages

  1. [1]

    Ahmadi and T

    M. Ahmadi and T. Farsadi, Multidisciplinary optimization of high aspect ratio composite wings with geometrical nonlinearity and aeroelastic tailoring. Aerospace Science and Technology, 2024. 145: p. 108849

    work page 2024

  2. [2]

    Z. Liu, J. Lu, and P. Zhu, Lightweight design of automotive composite bumper system using modified particle swarm optimizer. Composite Structures, 2016. 140: p. 630-643

    work page 2016

  3. [3]

    Yang and V.C

    E.- H. Yang and V.C. Li, Tailoring engineered cementitious composites for impact resistance. Cement and Concrete Research, 2012. 42(8): p. 1066-1071

    work page 2012

  4. [4]

    Al-Ketan, R.A

    O. Al-Ketan, R.A. Al-Rub, and R. Rowshan, Mechanical properties of a new type of architected interpenetrating phase composite materials. Adv. Mater. Technol, 2017. 2(2): p. 1600235

    work page 2017

  5. [5]

    Lacarbonara, M

    W. Lacarbonara, M. Talò, B. Carboni, and G. Lanzara. Tailoring of hysteresis across different material scales. in Recent Trends in Applied Nonlinear Mechanics and Physics: Selected Papers from CSNDD 2016. 2018. Springer

    work page 2016

  6. [6]

    Supian, S

    A. Supian, S. Sapuan, M. Zuhri, E. Zainudin, and H.H. Ya, Hybrid reinforced thermoset polymer composite in energy absorption tube application: A review. Defence Technology, 2018. 14(4): p. 291-305

    work page 2018

  7. [7]

    Jang, H.U

    K.-I. Jang, H.U. Chung, S. Xu, C.H. Lee, H. Luan, J. Jeong, H. Cheng, G.- T. Kim, S.Y. Han, and J.W. Lee, Soft network composite materials with deterministic and bio-inspired designs. Nature communications, 2015. 6(1): p. 6566

    work page 2015

  8. [8]

    Xie, G.P

    Y.M. Xie, G.P. Steven, Y. Xie, and G. Steven, Basic evolutionary structural optimization. 1997: Springer

    work page 1997

Show all 36 references
  1. [9]

    Bendsøe and N

    M.P. Bendsøe and N. Kikuchi, Generating optimal topologies in structural design using a homogenization method. Computer methods in applied mechanics and engineering, 1988. 71(2): p. 197-224

    1988

  2. [10]

    Rozvany, M

    G.I. Rozvany, M. Zhou, and T. Birker, Generalized shape optimization without homogenization. Structural optimization, 1992. 4: p. 250-252

    1992

  3. [11]

    T. Xie, X. Fu, O.-E. Ganea, R. Barzilay, and T. Jaakkola, Crystal diffusion variational autoencoder for periodic material generation. arXiv preprint arXiv:2110.06197, 2021

    2021 arXiv

  4. [12]

    J. Lim, S. Ryu, J.W. Kim, and W.Y. Kim, Molecular generative model based on conditional variational autoencoder for de novo molecular design. Journal of cheminformatics, 2018. 10: p. 1-9

    2018

  5. [13]

    Y. Dan, Y. Zhao, X. Li, S. Li, M. Hu, and J. Hu, Generative adversarial networks (GAN) based efficient sampling of chemical composition space for inverse design of inorganic materials. npj Computational Materials, 2020. 6(1): p. 84

    2020

  6. [14]

    Z. Yang, X. Li, L. Catherine Brinson, A.N. Choudhary, W. Chen, and A. Agrawal, Microstructural materials design via deep adversarial learning methodology. Journal of Mechanical Design, 2018. 140(11): p. 111416

    2018

  7. [15]

    Bastek and D.M

    J.-H. Bastek and D.M. Kochmann, Inverse design of nonlinear mechanical metamaterials via video denoising diffusion models. Nature Machine Intelligence, 2023. 5(12): p. 1466-1475

    2023

  8. [16]

    F. Sui, R. Guo, Z. Zhang, G.X. Gu, and L. Lin, Deep reinforcement learning for digital materials design. ACS Materials Letters, 2021. 3(10): p. 1433-1439

    2021

  9. [17]

    Brown, A.P

    N.K. Brown, A.P. Garland, G.M. Fadel, and G. Li, Deep reinforcement learning for engineering design through topology optimization of elementally discretized design domains. Materials & Design, 2022. 218: p. 110672

    2022

  10. [18]

    Brown, A.P

    N.K. Brown, A.P. Garland, G.M. Fadel, and G. Li, Deep reinforcement learning for the rapid on-demand design of mechanical metamaterials with targeted nonlinear deformation responses. Engineering Applications of Artificial Intelligence, 2023. 126: p. 106998

    2023

  11. [19]

    Elton, Z

    D.C. Elton, Z. Boukouvalas, M.D. Fuge, and P.W. Chung, Deep learning for molecular design—a review of the state of the art. Molecular Systems Design & Engineering, 2019. 4(4): p. 828-849

    2019

  12. [20]

    Bian and X.-Q

    Y. Bian and X.-Q. Xie, Generative chemistry: drug discovery with deep learning generative models. Journal of Molecular Modeling, 2021. 27: p. 1-18

    2021

  13. [21]

    L. Chen, W. Zhang, Z. Nie, S. Li, and F. Pan, Generative models for inverse design of inorganic solid materials. J. Mater. Inform, 2021. 1(4)

    2021

  14. [22]

    X. Wang, Y. Sheng, J. Ning, J. Xi, L. Xi, D. Qiu, J. Yang, and X. Ke, A critical review of machine learning techniques on thermoelectric materials. The Journal of Physical Chemistry Letters, 2023. 14(7): p. 1808-1822

    2023

  15. [23]

    S. Oh, Y. Jung, S. Kim, I. Lee, and N. Kang, Deep generative design: Integration of topology optimization and generative models. Journal of Mechanical Design, 2019. 141(11): p. 111405

    2019

  16. [24]

    Bessa, P

    M.A. Bessa, P. Glowacki, and M. Houlder, Bayesian machine learning in metamaterial design: Fragile becomes supercompressible. Advanced Materials, 2019. 31(48): p. 1904845

    2019

  17. [25]

    (Methods). The variance preserved form uses a scaling operation to el iminate the bias brought by the ground-truth value when adding noise and perform ing manipulation in the range of standard Gaussian distributions. However, it does not work well for properties that c annot s...

  18. [26]

    Song and S

    Y. Song and S. Ermon, Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 2019. 32

    2019

  19. [27]

    J. Ho, A. Jain, and P. Abbeel, Denoising diffusion probabilistic models. Advances in neural information processing systems, 2020. 33: p. 6840-6851

    2020

  20. [28]

    Yim, B.L

    J. Yim, B.L. Trippe, V. De Bortoli, E. Mathieu, A. Doucet, R. Barzilay, and T. Jaakk ola, Se (3) diffusion model with application to protein backbone generation. arXiv preprint arXiv:2302.02277, 2023

    2023 arXiv

  21. [29]

    Y. Song, J. Sohl- Dickstein, D.P. Kingma, A. Kumar, S. Ermon, and B. Poole, Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020

    2011 arXiv

  22. [30]

    Leach, S.M

    A. Leach, S.M. Schmon, M.T. Degiacomi, and C.G. Willcocks. Denoising diffusion probabilistic models on so (3) for rotational alignment. in ICLR 2022 Workshop on Geometrical and Topological Representation Learning. 2022

    2022

  23. [31]

    Guan, W.W

    J. Guan, W.W. Qian, X. Peng, Y. Su, J. Peng, and J. Ma, 3d equivariant diffusion for target- aware molecule generation and affinity prediction. arXiv preprint arXiv:2303.03543, 2023

    2023 arXiv

  24. [32]

    collision

    to synthesize the spatial representation of each fiber (Su pplementary Notes ). Overall, t he network for noise estimation used by the diffusion model consists of (1) a graph attention network (GAT) to encode each fiber’s 3D spatial representation, (2) a backbone of t ransform...

  25. [33]

    Ho and T

    J. Ho and T. Salimans, Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 32. ✁ ✂ ✄ ☎ ✆ ✝✞ ✟ ✠ ✆ ✡ ☛ ☞ ✁ ✌ ✍ ✎ ✍ ✏ ✍ ☎ ☎ ☛ ✑ ✁ ✌ ✒ ✓ ✒ ✔ ✟ ✠ ✒ ☛ ✑ ✁ ✕ ✟ ✖ ✄ ✏ ✟ ☛ ✁ ✗ ✆ ✟ ☛ ✒ ✔ ✘ ✙ ✁ ✚ ✄ ✔ ✛ ✆ ✟ ☛ Graph attention networks. arXiv preprint arXiv:1710.10903, 2017

    2022 arXiv

  26. [34]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L ✁ ✜ ✟ ✔ ✄ ✓ ☛ ✑ ✁ ✢ ✁ ☞ ✟ ✖ ✄ ✣ ☛ ✤ ✁ ✥ ✒ ✆ ✓ ✄ ✏ ☛ ✒ ✔ ✘ ✦ ✁ Polosukhin, Attention is all you need. Advances in neural information processing systems,

  27. [35]

    Savjolova, Preface to novye metody issledovanija tekstury polikristalliceskich materialov

    T. Savjolova, Preface to novye metody issledovanija tekstury polikristalliceskich materialov. Metallurgija, Moscow, 1985. 4(2): p. 6.2

    1985

  28. [36]

    !&,!%$+(%1$-!1;%(*/;!1;$!,18.-8%-! -$#&81&(.!&.!1;$!*.&1!()!-$/%$$,6! ! ! !!! !

    A. Kendall, Y. Gal, and R. Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. in Proceedings of the IEEE conference on computer vision and pattern recognition. 2018. Author contributions F.L., G.L . and S.P. conceived and designed ...

    2018

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.