Pith. sign in

REVIEW 4 major objections 5 minor 75 references

Rejoining fragmented ancient bamboo slips with physics-driven deep learning

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A physics-driven simulator of bamboo fracture and corrosion produces synthetic training pairs that raise Top-50 fragment-matching accuracy from 36% to 52% on a 1,350-candidate pool.

desk verdict A potentially useful benchmark and training recipe for bamboo slip fragment matching, but the headline Top-50 gain is undermined by an unstated calibration/test separation and a few underspecified modeling details. read the letter →

arxiv 2505.08601 v2 pith:3Q7WQPXC submitted 2025-05-13 cs.CV cond-mat.mtrl-sci

Jinchi Zhu , Zhou Zhao , Hailong Lei , Xiaoguang Wang , Jialiang Lu , Jing Li , Qianqian Tang , Jiachen Shen
show 3 more authors
Gui-Song Xia Bo Du Yongchao Xu
This is my paper · ORCID
classification cs.CVcond-mat.mtrl-sci
keywords bambooslipsfragmentrejoiningphysics-drivendeeplearningfracturemechanicscorrosionsimulationsynthetictrainingdatatripletnetworkcurvematching
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 claims that bamboo-slip rejoining, a task where manually paired training examples are almost impossible to produce, can be solved without any real paired data by training a matching network on synthetic pairs generated from a physical model of bamboo fracture and corrosion. The central numbers are Top-50 matching accuracy of 52.54% with 1,114 interference fragments present, against 36.02% for the leading scale-invariant curve-matching baseline. If the claim holds, it means a purely synthetic, physics-grounded training set transfers to real archaeological fragments well enough to be useful in practice, and that archaeologists using the system can verify suggested joins roughly twenty times faster than manual search. The paper also argues that the same recipe—model the material's fracture and degradation physics, synthesize pairs, train a ranker—can be transplanted to other fragmented artifacts.

What carries the argument

The load-bearing object is a two-stage simulator. First, fracture: the bamboo slip is discretized into vertical fiber bundles, and a Mode III (out-of-plane shear) stress field at each fracture front determines the probability density of the next fiber's fracture-angle change; through repeated sampling this produces a Markov-chain-style fracture curve. Second, corrosion: a multi-step erosion process reduces each fiber height in proportion to its geometric exposure, computed as $S_i = \mathrm{ReLU}(y_i - y_{i-1}) + \mathrm{ReLU}(y_i - y_{i+1})$, so protruding bundles erode faster. The simulator's parameters are tuned by a genetic algorithm that minimizes the separation between synthetic and real curve distributions after t-SNE projection, scored by silhouette. The trained network is a TripletNet that embeds 64-point normalized fracture curves and ranks candidates by learned similarity; at inference it returns Top-50 lists that archaeologists verify.

What would settle it

Compare the released fracture-curve data against the Bamboo236 fragments: if any of the 200 curves used to tune the simulator also appears among the 236 test fragments, the reported Top-k accuracies are inflated; if they are disjoint, rerun the pipeline with calibration and test sets drawn from different excavation sites to confirm the 52.54% result.

Watch

Extended reading notes

Core claim

WisePanda's core claim is that the forward process of bamboo-slip fragmentation can be simulated well enough to replace human-labeled training data. A probabilistic fracture engine, built on a Mode III (out-of-plane shear) stress field and a Markov-chain walk across vertical fiber bundles, generates initial breakage curves; an iterative corrosion model, using a ReLU-based exposure measure that erodes protruding fiber bundles faster, turns those curves into aged fragments. A triplet-trained matching network learns embeddings of 64-point normalized fracture curves from these synthetic pairs. On 118 expert-verified real pairs (Bamboo236) it reaches 94.07% Top-50 accuracy, and when 1,114 unrelated fragments are added (Bamboo1350) it keeps 52.54% Top-50 accuracy, versus 36.02% for the leading baseline SIS. On wooden slips, applied under the same protocol, it reaches 32.99% Top-50 accuracy on 335 expert-verified pairs, which the paper reads as evidence that the physics, not just the curve statistics, transfers.

Load-bearing premise

The simulator's parameters are tuned against a set of 200 real fracture curves, and the test fragments come from the same unpublished collection; the paper does not state that these two sets are disjoint, so the reported Top-k gains could partly come from tuning to the test distribution.

Editorial extensions

If this is right

  • On the 118-pair Bamboo236 set, WisePanda's Top-50 accuracy is 94.07%, so roughly nine of ten correct joins fall inside the shortlist an archaeologist would actually inspect.
  • With 1,114 interference fragments added (Bamboo1350), Top-50 accuracy remains 52.54%, more than 16 points above the best classical baseline, meaning the ranking advantage survives realistic candidate-pool sizes.
  • Because training pairs are synthesized from physics rather than manually matched, the method can be applied to newly excavated, unpublished bamboo-slip collections where labeled pairs do not exist.
  • The same physics-plus-synthesis recipe transfers to wooden slips, where WisePanda reaches 32.99% Top-50 accuracy on 335 expert-verified pairs, indicating the approach is not bamboo-specific.
  • Archaeologists using the system verified joins in roughly 15 seconds per query versus 326 seconds manually, a roughly 20-fold workflow speedup.

Reading between the lines

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

  • Beyond the paper: the most decisive check is a clean train/test split; if the 200 calibration curves are disjoint from the 236 test fragments, the 52.54% result survives, but if they overlap, the reported gains would be inflated by tuning the simulator to the test distribution.
  • Beyond the paper: the same two-stage recipe—fracture mechanics plus exposure-driven degradation—should be transferable to ceramic sherds, fresco fragments, and bone, where paired reassembly data is equally scarce but fracture physics is known; the paper sketches this extension but does not test it.
  • Beyond the paper: a useful stress test would withhold calibration curves from one excavation site and measure Top-50 accuracy on a different site, which would separate physical transfer from distribution fitting.
  • Beyond the paper: the reported 20x speedup measures assisted search time, not accuracy; in practice the speedup is only as good as the Top-50 recall plus the expert's ability to reject false candidates, so reporting precision at Top-50 alongside recall would make the archaeologist workflow easier to evaluate.
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 / 5 minor

Summary. WisePanda proposes a physics-driven deep learning framework for rejoining fragmented ancient bamboo slips. It simulates fracture propagation across bamboo fiber bundles under Mode III stress, models long-term geometric corrosion with an iterative ReLU-based exposure rule, and uses a genetic algorithm to tune the simulator parameters against 200 real fragment curves. Synthetic paired fracture curves are used to train a TripletNet-style matching network that ranks candidate matches for a query fragment. The authors report Top-k accuracies on 118 expert-verified bamboo pairs (Bamboo236), an expanded pool with 1,114 distractors (Bamboo1350), and wooden-slip datasets, with the headline result being 52.54% Top-50 accuracy on Bamboo1350 versus 36.02% for SIS. They also report a roughly 20-fold speedup in archaeologist workflow time using the accompanying software tool.

Significance. If the reported results hold, WisePanda is a practically valuable contribution to cultural-heritage restoration: it addresses a real bottleneck in bamboo-slip rejoining, releases code and curve-level datasets for independent reproduction, and evaluates against expert-verified ground truth rather than against its own fitted equations. The evaluation is thus not circular in the narrow derivation-to-fit sense. The cross-material evaluation on wooden slips is a further strength, and the open-source release is commendable. However, the significance is conditional on the calibration/test disjointness, the correctness of the fracture-angle sampler, and the absence of a control that isolates the physics component. These issues affect the central claim that physics-driven synthetic data is what produces the reported gains, rather than merely the network architecture or the volume of synthetic data.

major comments (4)
  1. [Methods, 'Parameter optimization'; Datasets; Extended Figure E3] The central claim that the physics simulator improves matching depends on the simulator being calibrated without using the 118 expert-verified test pairs. The manuscript states that parameter optimization uses a reference set of 200 real fragment curves (set A) but never states whether those 200 curves are disjoint from the 236 fragments in Bamboo236, and both are drawn from the same unpublished corpus. The concern is sharpened by Extended Figure E3b, which shows 'fracture matching quality matrices for 118 real fragment pairs' with the best matrix attributed to the most effective parameter setting, and by the Methods sentence that parameter optimization 'directly improves the model's ability to identify correct fragment matches.' If the calibration curves overlap the test fragments, or if the 118-pair matrices were used to select parameters, the reported 52.54% versus 36.02% Top-50 gap is an estimate of fit to the test distribution, not of generalization. The authors should state the disjointness explicitly and, if overlap exists, re-run the evaluation with a held-out split.
  2. [Methods, Eqs. (4)-(5) and Algorithm 1] The fracture-angle sampler is underspecified and algebraically inconsistent. Substituting Eq. (1) into Eq. (4) does not yield Eq. (5): the square root in the denominator of Eq. (4) does not disappear under the substitution, so the two expressions cannot both be correct. In addition, Eq. (5) is not a normalized probability density over Δθ and can be negative or undefined whenever cos(θi−1 + Δθ) ≤ 0, while no bounds on Δθ or a rejection/normalization scheme are provided for the SampleAngleChange routine in Algorithm 1. Since this sampler is the basis for all synthetic fracture curves used to train WisePanda, the physical data-generation pipeline is not fully specified as written and may generate invalid angle transitions.
  3. [Experimental evaluation, Results; Discussion] No ablation isolates the contribution of the physics-driven simulator. Table 1 compares WisePanda only against classical curve-matching methods (DTW, FMM, SIS) and manual random search; it does not compare the same TripletNet architecture trained on synthetic data generated without the fracture/corrosion model, on random curve perturbations, or on real curves alone. Without such a control, the abstract's statement that 'incorporating physical principles into deep learning models can significantly enhance their performance' is not directly supported: the improvement could come from the network architecture, the amount of synthetic data, or the triplet/ranking loss rather than from the physical model.
  4. [Table 1; Results and analysis] All reported accuracies are single-run values with no confidence intervals or multiple-seed variance. With only 118 query fragments, a single pair flip changes Top-k accuracy by about 0.85 percentage points, so the 16.5-point Top-50 gap is likely robust, but the absence of variance information makes it impossible to assess the stability of the smaller differences, such as the Top-1 results on Bamboo1350 where WisePanda (5.93%) is below SIS (7.20%) and FMM (6.36%). The authors should report standard deviations or confidence intervals over at least several seeds, and should soften the claim that WisePanda 'consistently outperforms' all baselines given these Top-1 exceptions.
minor comments (5)
  1. [Table 1 caption] The caption states that 'WisePanda consistently outperforms traditional curve matching approaches in both material types,' but the Top-1 rows for Bamboo1350 and Wood670 show WisePanda below SIS and FMM; the caption should be revised to describe the specific metrics where WisePanda leads.
  2. [Figure 3] There is a typo in the figure: 'Trainning data' should read 'Training data.'
  3. [Training details, Eq. (13)] Eq. (13) is described as a 'combined loss function,' but only one term is shown and the d_pos/d_neg notation is not defined; clarify how Eq. (13) relates to the triplet loss in Eq. (12).
  4. [Extended Figure E5] The 'approximately 20 times faster' claim compares WisePanda software search time with manual rejoining time across five cases; this is not a controlled comparison and should be described as an anecdotal workflow time, not a benchmark.
  5. [Abstract and main text] The abstract contains a subject-verb agreement issue ('Bamboo slips are a crucial medium ... and offers invaluable'), and the main text contains a sentence fragment: 'Training WisePanda presents a unique paradox: while manual fragment rejoining is prohibitively time-consuming. The very problem we aim to solve - this same process...'

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found: WisePanda's physics-driven pipeline is trained on synthetic data and scored on expert-verified real fragment pairs, so the headline accuracy is an independent empirical result.

full rationale

The claimed derivation chain is not circular. The fracture and corrosion simulator is an explicit probabilistic model (Eqs. 1-8), and its parameters are calibrated by a genetic algorithm against a distributional silhouette score on real fracture curves (Methods, 'Parameter optimization'), not against matching ground truth. Synthetic paired fragments from this simulator are then used to train a TripletNet, and the reported Top-k accuracies are measured on expert-verified real pairs (Bamboo236 and Wood670) against external baselines (DTW, FMM, SIS), so the central performance claim is empirically falsifiable rather than equivalent to the simulator's inputs. The matching-quality matrix shown in Extended Figure E3 for 118 real pairs is presented as confirmation/visualization of the optimized parameters, not as the genetic algorithm's fitness function, so it does not by itself establish a circular fit. The paper does not state whether the 200 calibration curves are disjoint from the 236 test fragments; if they overlap, that would be a data-hygiene or leakage concern for generalization claims, but it is not a derivation-to-fit circularity of the kind where a prediction reduces to a fitted parameter by construction. No load-bearing self-citation chain or imported uniqueness theorem appears; the fracture-mechanics citations are external and the modeling equations do not assume the target matching result. Therefore no circular step is demonstrated.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central empirical claim depends on three layers: a physics-inspired generator, a matching network, and an evaluation set. The generator introduces several fitted constants (c_e, N_c, stress and angle parameters, discretization) that are tuned by a genetic algorithm against a real-curve distribution, not measured. The model also assumes a specific fracture mode and a ReLU exposure rule for corrosion that are plausible but not independently validated. The evaluation assumes expert labels are correct and that calibration and test fragments are disjoint; the latter is not guaranteed in the text. No new physical entities are introduced.

free parameters (5)
  • Corrosion rate coefficient c_e = not reported (optimized by genetic algorithm)
    Controls how much exposed fiber height is removed per iteration in Eq. (8); tuned to match 200 real curves.
  • Corrosion iteration count N_c = not reported (optimized)
    Number of time steps in Algorithm 2; co-determines final degradation together with c_e.
  • Stress and angle sampling constants (V_xy, sigma, alpha, theta_init) = not reported
    Appear in the angle-change density Eq. (5); alpha evolves along the path and the constants are not measured, only argued to affect relative probabilities under an unnormalized density.
  • Fiber bundle width delta_x and count N_f = not reported
    Discretization of the slip into fiber bundles in Algorithm 1; chosen by hand rather than from microstructure data.
  • Genetic algorithm fitness weight lambda = not reported
    Balances silhouette minimization against diversity in Eq. (10); hand chosen.
assumptions (5)
  • domain assumption Bamboo slips are well approximated as N_f parallel fiber bundles of uniform width delta_x, and Mode III out-of-plane shear is the dominant burial stress.
    Invoked in Methods 'Physical model for data generation' to justify the Markov fracture model; no direct measurement of burial stress mode is provided.
  • ad hoc to paper Eq. (5) defines a valid sampling distribution for fracture angle changes.
    No normalization constant is given and cos(theta_{i-1}+Delta theta) can be negative, yet Algorithm 1 samples from it; the genetic algorithm fit can compensate only partially.
  • ad hoc to paper ReLU-exposure corrosion rule (Eqs. 7 and 8) captures long-term deterioration of bamboo edges.
    Erosion proportional to local protrusion is plausible but not validated against time-lapse material data.
  • domain assumption t-SNE plus silhouette score is a sufficient proxy for realism of simulated curves.
    Parameters are optimized to minimize overlap of 2D t-SNE embeddings; distortions introduced by the 2D embedding are not accounted for.
  • domain assumption Expert-verified fragment pairs are correct ground truth.
    Ground truth was produced by three archaeologists on unpublished artifacts; labels cannot be independently checked from released data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Rejoining fragmented ancient bamboo slips with physics-driven deep learning." pith.science (2026). https://pith.science/paper/3Q7WQPXC

@misc{pith2026250508601,
  author       = {Pith},
  title        = {Pith review of: Rejoining fragmented ancient bamboo slips with physics-driven deep learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3Q7WQPXC}},
  note         = {Machine review of arXiv:2505.08601}
}
read the original abstract

Bamboo slips are a crucial medium for recording ancient civilizations in East Asia, and offers invaluable archaeological insights for reconstructing the Silk Road, studying material culture exchanges, and global history. However, many excavated bamboo slips have been fragmented into thousands of irregular pieces, making their rejoining a vital yet challenging step for understanding their content. Here we introduce WisePanda, a physics-driven deep learning framework designed to rejoin fragmented bamboo slips. Based on the physics of fracture and material deterioration, WisePanda automatically generates synthetic training data that captures the physical properties of bamboo fragmentations. This approach enables the training of a matching network without requiring manually paired samples, providing ranked suggestions to facilitate the rejoining process. Compared to the leading curve matching method, WisePanda increases Top-50 matching accuracy from 36% to 52% among more than one thousand candidate fragments. Archaeologists using WisePanda have experienced substantial efficiency improvements (approximately 20 times faster) when rejoining fragmented bamboo slips. This research demonstrates that incorporating physical principles into deep learning models can significantly enhance their performance, transforming how archaeologists restore and study fragmented artifacts. WisePanda provides a new paradigm for addressing data scarcity in ancient artifact restoration through physics-driven machine learning.

Figures

Figures reproduced from arXiv: 2505.08601 by the authors.

Figure 1
Figure 1. The challenge of bamboo slip rejoining in archaeological research. Some excavated bamboo slip fragments from the Qin dynasty (221–206 BCE) arranged in a spiral pattern (left) illustrate the overwhelming scale of the rejoining challenge faced by archaeologists—with thousands of fragments at single excavation sites, each potentially matching with any other piece. A zoomed view of a repre￾sentative fragment (upper righ… view at source ↗
Figure 2
Figure 2. Physics-driven modeling of bamboo slip fracture and deterioration. a, The breakage process of bamboo slips showing how fracture propagates across the bamboo’s fiber structure, with the resulting irregular curve composed of black line segments and the corresponding stress field distribution (blue gradient). b, Detailed stress field model illustrating the mathematical relationship between the fracture angles (θi−1, θi… view at source ↗
Figure 3
Figure 3. Framework of WisePanda integrating physics-driven deep learning for bamboo slip rejoining. The workflow diagram illustrates the complete pipeline of our system. The process begins with input bamboo slip fragments (top row), where the target fragment is highlighted in pink. In the physical model stage, we implement two key components: a physics-driven fracture engine that gener￾ates initial breakage patterns, and a c… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

75 extracted references · 67 canonical work pages

  1. [1]

    Buried ideas: Legends of abdication and ideal government in early Chinese bamboo-slip manuscripts (State University of New York Press, 2015)

    Allan, S. Buried ideas: Legends of abdication and ideal government in early Chinese bamboo-slip manuscripts (State University of New York Press, 2015)

    work page 2015

  2. [2]

    Approaching the Tsinghua Bamboo Slips (Taylor & Francis, 2024)

    Liu, G. Approaching the Tsinghua Bamboo Slips (Taylor & Francis, 2024)

    work page 2024

  3. [3]

    From Bamboo Slips to Received Versions: Common Features in the Transformation of the” Laozi”

    Liu, X. From Bamboo Slips to Received Versions: Common Features in the Transformation of the” Laozi”. Harv. J. Asiat. Stud. 63, 337–382 (2003)

    work page 2003

  4. [4]

    Formulaic structure of Chu divinatory bamboo slips

    Ling, L. Formulaic structure of Chu divinatory bamboo slips. Early China 15, 71–86 (1990)

    work page 1990

  5. [5]

    A brief explanation of the Chu bamboo slips Hengxian

    Zheng, W. A brief explanation of the Chu bamboo slips Hengxian. Front. Philos. China 1, 418–431 (2006)

    work page 2006

  6. [6]

    Archaeological theory: an introduction (John Wiley & Sons, 2019)

    Johnson, M. Archaeological theory: an introduction (John Wiley & Sons, 2019)

    work page 2019

  7. [8]

    Shuihudi Qin Mu Zhujian (Shui hu di qin mu zhu jian) (ed Group, S. Q. T. B. S. C.) Title translates to: Bamboo Slips from the Qin Tomb at Shuihudi. [in Chinese] (Wenwu Chubanshe (Cultural Relics Publishing House), Beijing, 1990)

    work page 1990

  8. [9]

    Y., Lee, K

    Cha, M. Y., Lee, K. H. & Kim, Y. S. Micromorphological and chemical aspects of archaeological bamboos under long-term waterlogged condition. Int. Biodeterior. Biodegrad. 86, 115–121 (2014)

    work page 2014

Show all 75 references
  1. [10]

    & Hinton, G

    LeCun, Y., Bengio, Y. & Hinton, G. Deep learning. Nature 521, 436–444 (2015)

    2015

  2. [12]

    Silver, D. et al. Mastering the game of go without human knowledge. Nature 550, 354–359 (2017)

    2017

  3. [13]

    Senior, A. W. et al. Improved protein structure prediction using potentials from deep learning. Nature 577, 706–710 (2020)

    2020

  4. [14]

    Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021)

    2021

  5. [15]

    Dafoe, A. et al. Cooperative AI: machines must learn to find common ground. Nature 593, 33–36 (2021). 8

    2021

  6. [16]

    Merchant, A. et al. Scaling deep learning for materials discovery. Nature 624, 80–85 (2023)

    2023

  7. [17]

    Lu, M. Y. et al. A multimodal generative AI copilot for human pathology. Nature 634, 466–473 (2024)

    2024

  8. [18]

    Bodnar, C. et al. A foundation model for the Earth system. Nature, 1–8 (2025)

    2025

  9. [19]

    Bickler, S. H. Machine learning arrives in archaeology. Proc. of Adv. Archaeol. Pract. 9, 186–191 (2021)

    2021

  10. [20]

    & Yamazaki, K

    Meng, L., Kamitoku, N. & Yamazaki, K. Recognition of oracle bone inscriptions using deep learning based on data augmentation. in MetroArchaeo (2018), 33–38

    2018

  11. [21]

    Zhang, C. et al. Data-driven oracle bone rejoining: A dataset and practical self-supervised learning scheme. in Proc. ACM SIGKDD Int. Conf. Knowledge Discovery Data Mining (2022), 4482–4492

    2022

  12. [22]

    Zhu, L., Zhou, Z. & Hu, D. Globally consistent reconstruction of ripped-up documents. IEEE Trans. on Pattern Anal. and Mach. Intell. 30, 1–13 (2007)

    2007

  13. [24]

    Li, M.-Y. et al. Investigation into the deterioration process of archaeological bamboo strips of China from four different periods by chemical and anatomical analysis. Polym. Degrad. Stab. 109, 71–78 (2014)

    2014

  14. [25]

    & Pane, C

    Eslami, D., Di Angelo, L., Di Stefano, P. & Pane, C. Review of computer-based methods for ar- chaeological ceramic sherds reconstruction. Virtual Archaeol. Rev. 11, 34–49 (2020)

    2020

  15. [26]

    Hou, M. et al. Novel method for virtual restoration of cultural relics with complex geometric struc- ture based on multiscale spatial geometry. ISPRS Int. J. Geo-Inf. 7, 353 (2018)

    2018

  16. [28]

    Tan, T. et al. Mechanical properties of functionally graded hierarchical bamboo structures. Acta Biomater. 7, 3796–3803 (2011)

    2011

  17. [29]

    Behera, D. et al. A review on bamboo fiber reinforced composites and their potential applications. Emergent Mater. 8, 157–172 (2025)

    2025

  18. [30]

    & Wang, X

    Chen, Z., Ma, R., Du, Y. & Wang, X. State-of-the-art review on research and application of original bamboo-based composite components in structural engineering. in Structures 35 (2022), 1010–1029

    2022

  19. [36]

    Goldberg, D. E. Genetic Algorithms in Search, Optimization, and Machine Learning (Addison- Wesley, Reading, MA, 1989)

    1989

  20. [39]

    & Mikolajczyk, K

    Tian, Y., Barroso Laguna, A., Ng, T., Balntas, V. & Mikolajczyk, K. HyNet: Learning local de- scriptor with hybrid similarity measure and triplet loss. Proc. of Adv. Neural Inf. Process. Syst. 33, 7401–7412 (2020)

    2020

  21. [41]

    Brown, B. et al. Tools for virtual reassembly of fresco fragments.Int. J. Herit. Digit. Era 1, 313–329 (2012). 9

    2012

  22. [43]

    & Basri, R

    Frenkel, M. & Basri, R. Curve matching using the fast marching method. in Proc. Int. Worksh. on Energy Minimiz. Methods in Comput. Vis. and Pattern Recognit. (2003), 35–51

    2003

  23. [45]

    & Shimizu, H

    Amada, S., Ichikawa, Y., Munekata, T., Nagase, Y. & Shimizu, H. Fiber texture and mechanical graded structure of bamboo. Compos. B: Eng. 28, 13–20 (1997)

    1997

  24. [46]

    & Latella, B

    Low, I., Che, Z. & Latella, B. Mapping the structure, composition and mechanical properties of bamboo. J. Mater. Res. 21, 1969–1976 (2006)

    2006

  25. [47]

    Reconstruction of Early Chinese Bamboo and Wood Manuscripts: A Review (1900–2010)

    Staack, T. Reconstruction of Early Chinese Bamboo and Wood Manuscripts: A Review (1900–2010). CSMC Occas. Pap. 5 (2016)

    2016

  26. [48]

    Damage of bamboo and wooden materials based on linear elastic fracture mechanics in garden design

    Haiyan, T. Damage of bamboo and wooden materials based on linear elastic fracture mechanics in garden design. Fractur. Struct. Integr. 10, 472–480 (2016)

    2016

  27. [49]

    P., Kim, J

    Singh, A. P., Kim, J. S., M¨ oller, R., Chavan, R. R. & Kim, Y. S. The pivotal role of microscopy in unravelling the nature of microbial deterioration of waterlogged wood: A review. Forests 15, 889 (2024). Methods Previous work In recent years, various approaches have been pro...

    2024

  28. [50]

    Zhou, J. et al. Identifying designs from incomplete, fragmented cultural heritage objects by curve- pattern matching. J. Electron. Imaging 26, 011022–011022 (2017)

    2017

  29. [51]

    & Geng, G

    Zhang, Y., Li, K., Chen, X., Zhang, S. & Geng, G. A multi feature fusion method for reassembly of 3D cultural heritage artifacts. J. Cult. Herit. 33, 191–200 (2018)

    2018

  30. [52]

    R., Maringanti, H

    Das, B. R., Maringanti, H. B. & Dash, N. S. in Digital. Cult. Technol. 92–97 (Routledge, 2022)

    2022

  31. [53]

    Artificial Intelligence In The Protection Of Intangible Cultural Heritage

    Tsatsanashvili, A. Artificial Intelligence In The Protection Of Intangible Cultural Heritage. Eur. J. Turk. Stud. 12, 163–178 (2024)

    2024

  32. [54]

    Tsesmelis, T. et al. Re-assembling the past: The RePAIR dataset and benchmark for real world 2D and 3D puzzle solving. Proc. of Adv. Neural Inf. Process. Syst. 37, 30076–30105 (2024)

    2024

  33. [55]

    Assael, Y. et al. Restoring and attributing ancient texts using deep neural networks. Nature 603, 280–283 (2022)

    2022

  34. [56]

    & Weng, Y

    Zheng, Y., Li, X. & Weng, Y. Reunion helper: an edge matcher for sibling fragment identification of the Dunhuang manuscript. Heritage Science 12, 52 (2024)

    2024

  35. [57]

    Zhang, Y. et al. LLMCO4MR: LLMs-Aided Neural Combinatorial Optimization for Ancient Manuscript Restoration from Fragments with Case Studies on Dunhuang. in Proc. Eur. Conf. on Comput. Vis. (2024), 253–269

    2024

  36. [58]

    Li, B. et al. Oracle Bone Inscriptions Multi-modal Dataset. arXiv preprint arXiv:2407.03900 (2024)

    2024 arXiv

  37. [59]

    Bindings of Ancient Chinese Bamboo and Wood Scrolls

    Staack, T. Bindings of Ancient Chinese Bamboo and Wood Scrolls. Stud. Manusc. Cult., 39 (2023)

    2023

  38. [60]

    Written on Bamboo and Silk: The Beginning of Chinese Books and Inscriptions 1964

    Chang, C.-s. Written on Bamboo and Silk: The Beginning of Chinese Books and Inscriptions 1964

    1964

  39. [61]

    Science and Civilisation in China: Volume 5, Chemistry and Chemical Technology, Part 5, Spagyrical Discovery and Invention: Physiological Alchemy (Cambridge University Press, 1983)

    Needham, J. Science and Civilisation in China: Volume 5, Chemistry and Chemical Technology, Part 5, Spagyrical Discovery and Invention: Physiological Alchemy (Cambridge University Press, 1983)

    1983

  40. [62]

    & Wang, G

    Luo, Z. & Wang, G. Liusha Zhuijian (Liu sha zhui jian) (Shangyu Luo shi yin xing, 1914)

    1914

  41. [63]

    & Hutson, J

    Ghaith, K. & Hutson, J. A qualitative study on the integration of artificial intelligence in cultural heritage conservation. Metaverse 5 (2024)

    2024

  42. [64]

    S., Heider, Y

    Aldakheel, F., Elsayed, E. S., Heider, Y. & Weeger, O. Physics-based machine learning for compu- tational fracture mechanics. Mach. Learn. Comput. Sci. Eng. 1, 18 (2025)

    2025

  43. [65]

    Introduction to computational materials science: fundamentals to applications (Cambridge University Press, 2013)

    LeSar, R. Introduction to computational materials science: fundamentals to applications (Cambridge University Press, 2013)

    2013

  44. [66]

    Investigation of the transverse fracture mechanisms of bamboo by the finite element method

    Ramful, R. Investigation of the transverse fracture mechanisms of bamboo by the finite element method. J. Mater. Sci. 57, 6233–6248 (2022)

    2022

  45. [67]

    Dixon, P. G. & Gibson, L. J. The structure and mechanics of Moso bamboo material. J. R. Soc. Interface 11, 20140321 (2014)

    2014

  46. [68]

    Chen, M. et al. Fracture modes of bamboo fiber bundles in three-point bending. Cellulose 26, 8101–8108 (2019)

    2019

  47. [69]

    & Sakuma, A

    Ramful, R. & Sakuma, A. Investigation of the effect of inhomogeneous material on the fracture mechanisms of bamboo by finite element method. Materials 13, 5039 (2020)

    2020

  48. [70]

    & Fei, B

    Huang, Y. & Fei, B. Comparison of the mechanical characteristics of fibers and cell walls from moso bamboo and wood. BioResources 12, 8230–8239 (2017)

    2017

  49. [71]

    Gdoutos, E. E. Fracture mechanics: an introduction (Springer Nature, 2020)

    2020

  50. [72]

    (US), F. P. L. Wood handbook: wood as an engineering material 72 (The Laboratory, 1987)

    1987

  51. [73]

    A., Haight, J

    Blanchette, R. A., Haight, J. E., Koestler, R. J., Hatchfield, P. B. & Arnold, D. Assessment of deterioration in archaeological wood from ancient Egypt. J. Am. Inst. Conserv. 33, 55–70 (1994)

    1994

  52. [74]

    Liu, L. et al. Cracks Repairing and Resistance to Water Penetration Properties of Microbial Self- Healing Cement. Eng. Life. Sci. 25, e70010 (2025)

    2025

  53. [75]

    & Hinton, G

    Van der Maaten, L. & Hinton, G. Visualizing data using t-SNE. J. Mach. Learn. Res. 9 (2008)

    2008

  54. [76]

    Goldberg, D. E. Genetic Algorithms in Search, Optimization, and Machine Learning (Addison- Wesley, Reading, MA, 1989). 18

    1989

  55. [77]

    Anderson, T. L. & Anderson, T. L. Fracture mechanics: fundamentals and applications (CRC press, 2005)

    2005

  56. [78]

    Hussain, A., Sakhaei, A. H. & Shafiee, M. Machine learning-based constitutive modelling for mate- rial non-linearity: A review. Mech. Adv. Mater. Struct., 1–19 (2024)

    2024

  57. [79]

    Kutz, J. N. Machine learning for parameter estimation. Proc. Natl. Acad. Sci. U. S. A. 120, e2300990120 (2023)

    2023

  58. [80]

    & O’Hagan, A

    Brynjarsdottir, J. & O’Hagan, A. Learning about physical parameters: The importance of model discrepancy. Inverse Probl. 30, 114007 (2014)

    2014

  59. [81]

    Detecting and quantifying ongoing decay of organic archaeological remains: A dis- cussion of different approaches

    Matthiesen, H. Detecting and quantifying ongoing decay of organic archaeological remains: A dis- cussion of different approaches. Quaternary International 368, 43–50 (2015)

    2015

  60. [82]

    Rousseeuw, P. J. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. J. Comput. Appl. Math. 20, 53–65 (1987)

    1987

  61. [83]

    & Ailon, N

    Hoffer, E. & Ailon, N. Deep metric learning using triplet network. in Proc. Int. Worksh. on Simil.- Based Pattern Recognit. (2015), 84–92

    2015

  62. [84]

    Funkhouser, T. et al. Learning how to match fresco fragments. ACM J. Comput. Cult. Herit. 4, 1–13 (2011)

    2011

  63. [85]

    & Guardiani, E

    Di Angelo, L., Di Stefano, P. & Guardiani, E. A review of computer-based methods for classification and reconstruction of 3D high-density scanned archaeological pottery. J. Cult. Herit. 56, 10–24 (2022)

    2022

  64. [86]

    Smith, L. N. & Topin, N. Super-convergence: Very fast training of neural networks using large learning rates. in AI ML Multi-Domain Oper. Appl. 11006 (2019), 369–386

    2019

  65. [87]

    Berndt, D. J. & Clifford, J. Using dynamic time warping to find patterns in time series. in Proc. ACM SIGKDD Int. Conf. Knowl. Discov. Data Min. (1994), 359–370

    1994

  66. [88]

    J., Shakiban, C., Tannenbaum, A

    Calabi, E., Olver, P. J., Shakiban, C., Tannenbaum, A. & Haker, S. Differential and numerically invariant signature curves applied to object recognition. Int. J. Comput. Vis. 26, 107–135 (1998)

    1998

  67. [89]

    & Basri, R

    Frenkel, M. & Basri, R. Curve matching using the fast marching method. in Proc. Int. Worksh. on Energy Minimiz. Methods in Comput. Vis. and Pattern Recognit. (2003), 35–51. 19 Extended Figure a b Figure E1: Typical fracture patterns of ancient bamboo slips. (a), The upper row ...

    2003

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.