pith. machine review for the scientific record. sign in

arxiv: 2605.07521 · v1 · submitted 2026-05-08 · 💻 cs.AI

Recognition: no theorem link

From Feasible to Practical: Pareto-Optimal Synthesis Planning

Friedrich Hastedt , Dongda Zhang , Antonio del Rio Chanona
Authors on Pith no claims yet

Pith reviewed 2026-05-11 01:48 UTC · model grok-4.3

classification 💻 cs.AI
keywords Pareto frontmulti-objective optimizationretrosynthesissynthesis planningA* searchweighted scalarizationroute optimizationtrade-off analysis
0
0 comments X

The pith

MORetro* recovers the true Pareto front of synthesis routes for any fixed single-step retrosynthesis model.

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

Standard computer-aided synthesis planning stops once it finds one workable route, using metrics like length or convergence. This paper reframes the task as a multi-objective search that must balance several competing goals such as cost, toxicity, and yield at the same time. MORetro* solves the problem by converting the objectives into weighted single-objective searches, guiding the exploration with informed sampling, and running a multi-objective A* procedure that carries optimality guarantees. A sympathetic reader would care because the output is no longer a single path but an explicit menu of non-dominated routes that lets a chemist choose according to their actual priorities rather than accepting whatever the algorithm found first.

Core claim

MORetro* formulates retrosynthesis as a multi-objective optimization task and uses weighted scalarization together with Bayesian-optimization-informed sampling inside a multi-objective A* framework. For any fixed single-step retrosynthesis model the algorithm recovers the complete Pareto front of synthesis routes, with formal optimality guarantees. On standard benchmarks the resulting fronts contain diverse, high-quality solutions that single-objective planners overlook.

What carries the argument

MORetro*, the algorithm that extends multi-objective A* search through weighted scalarization and BO-informed sampling to enumerate all non-dominated synthesis routes.

If this is right

  • Planning software can return explicit sets of routes showing cost-yield-sustainability trade-offs instead of a single path.
  • A chemist can select a route that matches their current priorities without needing to rerun the search under new weights.
  • Routes that look inferior under one metric alone can still be optimal when all criteria are considered together.
  • Any single-objective shortest-path method is guaranteed to miss at least some members of the true Pareto front.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the single-step model is allowed to improve during the search, the current optimality guarantees would need to be relaxed or replaced with online updating rules.
  • Non-convex portions of the front may still require additional techniques beyond scalarization to ensure full coverage.
  • The Pareto sets could serve as training data for learning improved single-step models that better predict realistic trade-offs.

Load-bearing premise

The single-step retrosynthesis model stays fixed and its predictions remain accurate enough that the computed trade-offs reflect real chemical behavior.

What would settle it

Run MORetro* on a small target molecule whose complete set of routes can be enumerated exhaustively and check whether every non-dominated route appears in the reported front and no dominated route does.

Figures

Figures reproduced from arXiv: 2605.07521 by Antonio del Rio Chanona, Dongda Zhang, Friedrich Hastedt.

Figure 1
Figure 1. Figure 1: Exemplary Pareto front generated by MORetro∗ . Each red point corresponds to a Pareto-optimal synthesis route given an arbitrary number of user-specified objectives. 1). MORetro∗ extends the single-objective Retro∗ algo￾rithm (Chen et al., 2020) to the multi-objective setting by down-sampling the high-dimensional objective space via linear (weight-based) scalarization. To effectively explore this space, we… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of MORetro∗ . During each iteration, multiple frontier nodes are picked according to different weight vectors, expanded and new solutions are recorded. After wbudget (M) iterations, weight vectors are re-sampled. estimates the minimum cost of synthesizing the target product t along routes that pass through m. We further introduce the reaction number rn(m|G) and the graph￾independent heuristic esti… view at source ↗
Figure 3
Figure 3. Figure 3: BO-informed sampling. The selected weights (drawn from Sobol sequences) for the next iteration are encircled. Larger dots show weights that were explored during the warm-up period. 5 [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Per-molecule comparison of Pareto front metrics for the ChEMBL (G2E) experiment. Statistics are normalized according to the best value found per molecule. line Dominance (%) is the fraction of baseline solutions dominated by MORetro∗ , and Self Dominance (%) is the fraction of MORetro∗ solutions dominated by the baseline. To address (1), [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Per-molecule normalized number of unique solutions for all datasets investigated. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
read the original abstract

Current computer-aided synthesis planning (CASP) methods often treat retrosynthesis as solved once a single feasible route is identified, focusing primarily on convergence or shortest-path metrics. This view is misaligned with real-world practice, where chemists must balance competing objectives such as cost, sustainability, toxicity, and overall yield. To address this, we formulate synthesis planning as a multi-objective search problem and introduce MORetro*, an algorithm that generates a Pareto front of synthesis routes to explicitly capture trade-offs among user-defined criteria. MORetro* uses weighted scalarization and BO-informed sampling to efficiently navigate the combinatorial search space and prioritize promising trade-offs. Building on multi-objective A*-search, we provide optimality guarantees showing that, for a fixed single-step model, MORetro* recovers the true Pareto front. Across multiple retrosynthesis benchmarks, MORetro* produces diverse, high-quality Pareto fronts, uncovering solutions overlooked by single-objective approaches and better aligning CASP outputs with industrial decision-making.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces MORetro*, a multi-objective retrosynthesis algorithm that formulates synthesis planning as a search over trade-offs among objectives such as cost, yield, sustainability, and toxicity. It employs weighted scalarization combined with Bayesian optimization-informed sampling inside a multi-objective A* framework, and claims optimality guarantees that, for any fixed single-step retrosynthesis model, the algorithm recovers the true Pareto front. Evaluations on standard retrosynthesis benchmarks are reported to produce more diverse and higher-quality route sets than single-objective baselines.

Significance. If the optimality guarantees are valid and the approach scales, the work would advance CASP by shifting focus from single feasible routes to explicit Pareto fronts that better match industrial multi-criteria decision making. The explicit provision of guarantees for a fixed single-step model is a constructive strength that distinguishes the contribution from purely heuristic multi-objective search methods.

major comments (2)
  1. [Abstract / MORetro* algorithm description] Abstract and method description: The central claim states that MORetro* recovers the true Pareto front via multi-objective A*-search optimality guarantees. However, the described procedure relies on weighted scalarization plus BO-informed sampling. Standard weighted-sum scalarization identifies only supported solutions on the convex hull and systematically misses non-supported points on non-convex fronts, which are expected for typical synthesis objectives (cost vs. toxicity, yield vs. sustainability). The manuscript must clarify whether epsilon-constraint methods, explicit Pareto-set maintenance, or exhaustive weight enumeration are used to restore completeness; without this, the optimality guarantee does not hold for the full front.
  2. [Abstract / Optimality guarantees section] Optimality guarantees paragraph: The abstract asserts that the guarantees apply for a fixed single-step model, yet no derivation, proof sketch, or reference to the specific multi-objective A* properties (e.g., admissible heuristic conditions or dominance pruning rules) is supplied. This absence prevents verification that the guarantee survives the scalarization step and is load-bearing for the paper's primary contribution.
minor comments (2)
  1. [Experimental evaluation] Benchmarks section: Provide explicit rules for data exclusion, error analysis of the single-step model predictions, and the exact procedure for weight sampling to support reproducibility.
  2. [Preliminaries] Notation: Define the Pareto front formally in the discrete space of synthesis routes and clarify how non-dominated routes are maintained during search.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address the major comments point by point below, with planned revisions to the manuscript.

read point-by-point responses

  1. Referee: [Abstract / MORetro* algorithm description] Abstract and method description: The central claim states that MORetro* recovers the true Pareto front via multi-objective A*-search optimality guarantees. However, the described procedure relies on weighted scalarization plus BO-informed sampling. Standard weighted-sum scalarization identifies only supported solutions on the convex hull and systematically misses non-supported points on non-convex fronts, which are expected for typical synthesis objectives (cost vs. toxicity, yield vs. sustainability). The manuscript must clarify whether epsilon-constraint methods, explicit Pareto-set maintenance, or exhaustive weight enumeration are used to restore completeness; without this, the optimality guarantee does not hold for the full front.

    Authors: We agree that standard weighted-sum scalarization recovers only supported solutions on the convex hull and can miss non-supported points on non-convex fronts. The MORetro* implementation maintains an explicit set of non-dominated solutions via dominance pruning within the multi-objective A* search and uses BO-informed sampling over weight vectors to explore trade-offs. However, this combination does not guarantee recovery of the complete Pareto front. We will revise the abstract and algorithm description to state that the optimality guarantees apply specifically to the supported solutions obtained from the scalarized subproblems, and we will add a discussion of the limitations for non-convex fronts along with the practical utility of the supported front for synthesis planning. revision: yes

  2. Referee: [Abstract / Optimality guarantees section] Optimality guarantees paragraph: The abstract asserts that the guarantees apply for a fixed single-step model, yet no derivation, proof sketch, or reference to the specific multi-objective A* properties (e.g., admissible heuristic conditions or dominance pruning rules) is supplied. This absence prevents verification that the guarantee survives the scalarization step and is load-bearing for the paper's primary contribution.

    Authors: We acknowledge that the manuscript currently lacks a derivation or proof sketch. In the revision we will add a dedicated subsection (or appendix) providing a proof sketch. The argument will rely on the standard admissibility conditions for multi-objective A* heuristics, the correctness of dominance pruning for preserving the Pareto set, and the fact that each scalarized single-objective A* search returns an optimal route for its weight vector when the single-step model is fixed. We will also include references to the multi-objective heuristic search literature to support the reasoning. revision: yes

Circularity Check

0 steps flagged

No significant circularity: optimality guarantee conditioned on external fixed model

full rationale

The paper's derivation chain centers on formulating synthesis planning as multi-objective search and invoking multi-objective A*-search to supply optimality guarantees that MORetro* recovers the true Pareto front for any fixed single-step retrosynthesis model. This guarantee is explicitly conditioned on the single-step model being an external, fixed input rather than something derived or fitted inside the method. Weighted scalarization and BO-informed sampling are presented as algorithmic choices for navigating the space, but the optimality claim is not shown to reduce to those choices by construction; it is instead asserted to inherit from the underlying A* framework. No self-definitional loops, fitted inputs renamed as predictions, or load-bearing self-citations appear in the provided abstract or described structure. The derivation therefore remains self-contained against external benchmarks and does not collapse to its own inputs.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach relies on standard multi-objective optimization primitives (weighted scalarization, A* search, Bayesian optimization sampling) and user-specified objective functions; no new physical entities or ad-hoc constants are introduced in the abstract.

free parameters (1)
  • objective weights
    User-chosen weights for scalarizing multiple criteria; these are inputs rather than fitted values but directly affect the recovered front.
axioms (1)
  • domain assumption Single-step retrosynthesis model is fixed and accurate
    Optimality guarantee is conditioned on this fixed model; invoked in the guarantee statement.

pith-pipeline@v0.9.0 · 5464 in / 1281 out tokens · 35082 ms · 2026-05-11T01:48:05.362093+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

48 extracted references · 48 canonical work pages

  1. [1]

    The 37th International Conference on Machine Learning (ICML) , year=

    Retro*: Learning Retrosynthetic Planning with Neural Guided A* Search , author=. The 37th International Conference on Machine Learning (ICML) , year=

    work page

  2. [2]

    Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages =

    Xie, Shufang and Yan, Rui and Han, Peng and Xia, Yingce and Wu, Lijun and Guo, Chenjuan and Yang, Bin and Qin, Tao , title =. Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages =. 2022 , doi =

    work page 2022

  3. [3]

    doi:10.1186/s13321-020-00472-1 , pages =

    Journal of Cheminformatics , author =. doi:10.1186/s13321-020-00472-1 , pages =

    work page doi:10.1186/s13321-020-00472-1

  4. [4]

    and Fong, Mun Hong and Roh, Jihye and Levin, Itai and Yu, Kevin and Joung, Joonyoung F

    Tu, Zhengkai and Choure, Sourabh J. and Fong, Mun Hong and Roh, Jihye and Levin, Itai and Yu, Kevin and Joung, Joonyoung F. and Morgan, Nathan and Li, Shih-Cheng and Sun, Xiaoqi and Lin, Huiqian and Murnin, Mark and Liles, Jordan P. and Struble, Thomas J. and Fortunato, Michael E. and Liu, Mengjie and Green, William H. and Jensen, Klavs F. and Coley, Conn...

    work page 2025

  5. [5]

    Segler and Mark P

    Marwin H.S. Segler and Mark P. Waller , doi =. Neural-Symbolic Machine Learning for Retrosynthesis and Reaction Prediction , volume =. Chemistry – A European Journal , pages =

    work page

  6. [6]

    Segler, Marwin H. S. and Waller, Mark P. , year =. Planning Chemical Syntheses with Deep Neural Networks and Symbolic. Nature , volume =

    work page

  7. [7]

    2024 , journal =

    An Algorithmic Framework for Synthetic Cost-Aware Decision Making in Molecular Design , author =. 2024 , journal =

    work page 2024

  8. [8]

    Chemical Science , author =

    Data-driven recommendation of agents, temperature, and equivalence ratios for organic synthesis , doi =. Chemical Science , author =

    work page

  9. [9]

    Chen, Shuan and Jung, Yousung , year =. Deep. JACS Au , volume =

    work page

  10. [10]

    Retrosynthesis Prediction with Conditional Graph Logic Network , booktitle =

    Dai, Hanjun and Li, Chengtao and Coley, Connor and Dai, Bo and Song, Le , year =. Retrosynthesis Prediction with Conditional Graph Logic Network , booktitle =

    work page

  11. [11]

    Computer‐assisted synthetic planning: The end of the beginning

    Szymku \'c , Sara and Gajewska, Ewa P and Klucznik, Tomasz and Molga, Karol and Dittwald, Piotr and Startek, Micha and Bajczyk, Micha and Grzybowski, Bartosz A. Computer‐assisted synthetic planning: The end of the beginning. Angew. Chem. Int. Ed Engl

    work page

  12. [12]

    Nature Communications , author =

    Retrosynthesis prediction using an end-to-end graph generative architecture for molecular graph editing , volume =. Nature Communications , author =. doi:10.1038/s41467-023-38851-5 , pages =

    work page doi:10.1038/s41467-023-38851-5

  13. [13]

    Molecule edit graph attention network: Modeling chemical reactions as sequences of graph edits

    Molecule Edit Graph Attention Network: Modeling Chemical Reactions as Sequences of Graph Edits , volume =. Journal of Chemical Information and Modeling , author =. doi:10.1021/acs.jcim.1c00537 , pages =

    work page doi:10.1021/acs.jcim.1c00537

  14. [14]

    Thirty-Fifth Conference on Neural Information Processing Systems , year=

    Learning Graph Models for Retrosynthesis Prediction , author=. Thirty-Fifth Conference on Neural Information Processing Systems , year=

    work page

  15. [15]

    and Bekas, Costas and Lee, Alpha A

    Schwaller, Philippe and Laino, Teodoro and Gaudin, Théophile and Bolgar, Peter and Hunter, Christopher A. and Bekas, Costas and Lee, Alpha A. , year =. Molecular Transformer: A Model for Uncertainty-Calibrated Chemical Reaction Prediction , volume =. doi:10.1021/acscentsci.9b00576 , pages =

    work page doi:10.1021/acscentsci.9b00576

  16. [16]

    , year =

    Karpov, Pavel and Godin, Guillaume and Tetko, Igor V. , year =. A Transformer Model for Retrosynthesis , doi =. Artificial Neural Networks and Machine Learning –

    work page

  17. [17]

    Advances in

    Yu, Yemin and Wei, Ying and Kuang, Kun and Huang, Zhengxing and Yao, Huaxiu and Wu, Fei , year =. Advances in

    work page

  18. [18]

    Retrosynthetic Planning with Dual Value Networks , booktitle =

    Liu, Guoqing and Xue, Di and Xie, Shufang and Xia, Yingce and Tripp, Austin and Maziarz, Krzysztof and Segler, Marwin and Qin, Tao and Zhang, Zongzhang and Liu, Tie-Yan , year =. Retrosynthetic Planning with Dual Value Networks , booktitle =

    work page

  19. [19]

    and Coley, Connor W

    Schreck, John S. and Coley, Connor W. and Bishop, Kyle J. M. , year =. Learning. ACS Central Science , volume =

    work page

  20. [20]

    Retrosynthesis Planning via Worst-path Policy Optimisation in Tree-structured

    Mianchu Wang and Giovanni Montana , booktitle=. Retrosynthesis Planning via Worst-path Policy Optimisation in Tree-structured

    work page

  21. [21]

    Advances in Neural Information Processing Systems , author =

    Depth-First Proof-Number Search with Heuristic Edge Cost and Application to Chemical Synthesis Planning , volume =. Advances in Neural Information Processing Systems , author =

    work page

  22. [22]

    Self-Improved Retrosynthetic Planning

    Junsu Kim and Sungsoo Ahn and Hankook Lee and Jinwoo Shin. Self-Improved Retrosynthetic Planning. Proceedings of the 38th International Conference on Machine Learning (ICML). 2021

    work page 2021

  23. [23]

    Retro-fallback: retrosynthetic planning in an uncertain world , year =

    Tripp, Austin and Maziarz, Krzysztof and Lewis, Sarah and Segler, Marwin and Hern\'. Retro-fallback: retrosynthetic planning in an uncertain world , year =. International Conference on Representation Learning (ICLR) , pages =

    work page

  24. [24]

    Communications Chemistry , author =

    Efficient retrosynthetic planning with. Communications Chemistry , author =. doi:10.1038/s42004-024-01133-2 , pages =

    work page doi:10.1038/s42004-024-01133-2

  25. [25]

    Tango*: constrained synthesis planning using chemically informed value functions

    Armstrong, Daniel and Jončev, Zlatko and Guo, Jeff and Schwaller, Philippe. Tango*: constrained synthesis planning using chemically informed value functions. Digital Discovery. 2025. doi:10.1039/D5DD00130G

    work page doi:10.1039/d5dd00130g 2025

  26. [26]

    , title =

    Yu, Kevin and Roh, Jihye and Li, Ziang and Gao, Wenhao and Wang, Runzhong and Coley, Connor W. , title =. 2024 , booktitle =

    work page 2024

  27. [27]

    and Saigiridharan, Lakshidaa and Genheden, Samuel

    Westerlund, Annie M. and Saigiridharan, Lakshidaa and Genheden, Samuel. Human-guided synthesis planning via prompting. Chemical Science. 2025. doi:10.1039/D5SC00927H

    work page doi:10.1039/d5sc00927h 2025

  28. [28]

    Vaucher, Andrea Byekwaso, Philippe Schwaller, Alessandra Toniato, and Teodoro Laino

    Thakkar, Amol and Vaucher, Alain C. and Byekwaso, Andrea and Schwaller, Philippe and Toniato, Alessandra and Laino, Teodoro , year =. Unbiasing Retrosynthesis Language Models with Disconnection Prompts , volume =. doi:10.1021/acscentsci.3c00372 , pages =

    work page doi:10.1021/acscentsci.3c00372

  29. [29]

    Multistep retrosynthesis combining a disconnection aware triple transformer loop with a route penalty score guided tree search , volume =

    Multistep retrosynthesis combining a disconnection aware triple transformer loop with a route penalty score guided tree search , volume =. Chemical Science , author =. doi:10.1039/D3SC01604H , pages =

    work page doi:10.1039/d3sc01604h

  30. [30]

    and Jončev, Zlatko and Schwaller, Philippe , year =

    Xuan-Vu, Nguyen and Armstrong, Daniel and Wehrbach, Milena and Bran, Andres M. and Jončev, Zlatko and Schwaller, Philippe , year =. Synthelite: Chemist-aligned and feasibility-aware synthesis planning with. doi:10.48550/arXiv.2512.16424 , publisher =

    work page doi:10.48550/arxiv.2512.16424

  31. [31]

    and Clevert, Djork-Arné and Preuss, Mike and Genheden, Samuel , year =

    Lai, Helen and Kannas, Christos and Hassen, Alan Kai and Granqvist, Emma and Westerlund, Annie M. and Clevert, Djork-Arné and Preuss, Mike and Genheden, Samuel , year =. Multi-objective synthesis planning by means of Monte Carlo Tree search , volume =. doi:10.1016/j.ailsci.2025.100130 , journal =

    work page doi:10.1016/j.ailsci.2025.100130 2025

  32. [32]

    ChemRxiv , year =

    Alan Kai Hassen and Helen Lai and Samuel Genheden and Mike Preuss and Djork-Arné Clevert , title =. ChemRxiv , year =

    work page

  33. [34]

    doi:10.1021/acs.jcim.0c01426 , pages =

    Journal of Chemical Information and Modeling , author =. doi:10.1021/acs.jcim.0c01426 , pages =

    work page doi:10.1021/acs.jcim.0c01426

  34. [35]

    Chemical Society Reviews , author =

    Chemical data intelligence for sustainable chemistry , volume =. Chemical Society Reviews , author =. doi:10.1039/D1CS00477H , pages =

    work page doi:10.1039/d1cs00477h

  35. [36]

    doi:10.1186/s40360-018-0282-6 , pages =

    Pu, Limeng and Naderi, Misagh and Liu, Tairan and Wu, Hsiao-Chun and Mukhopadhyay, Supratik and Brylinski, Michal , year =. doi:10.1186/s40360-018-0282-6 , pages =

    work page doi:10.1186/s40360-018-0282-6

  36. [37]

    doi:10.1186/s13321-025-01076-3 , pages =

    Journal of Cheminformatics , author =. doi:10.1186/s13321-025-01076-3 , pages =

    work page doi:10.1186/s13321-025-01076-3

  37. [38]

    Chemical Science , author =

    Towards efficient discovery of green synthetic pathways with Monte Carlo tree search and reinforcement learning , volume =. Chemical Science , author =. doi:10.1039/D0SC04184J , pages =

    work page doi:10.1039/d0sc04184j

  38. [39]

    and Coley, Connor W

    Gao, Hanyu and Struble, Thomas J. and Coley, Connor W. and Wang, Yuran and Green, William H. and Jensen, Klavs F. , year =. Using Machine Learning To Predict Suitable Conditions for Organic Reactions , volume =. doi:10.1021/acscentsci.8b00357 , pages =

    work page doi:10.1021/acscentsci.8b00357

  39. [40]

    and Leslie, David S

    Moss, Henry B. and Leslie, David S. and Gonz\'. GIBBON: general-purpose information-based Bayesian optimisation , year =. J. Mach. Learn. Res. , articleno =

    work page

  40. [41]

    and Schuffenhauer, A

    Estimation of synthetic accessibility score of drug-like molecules based on molecular complexity and fragment contributions , volume =. Journal of Cheminformatics , author =. doi:10.1186/1758-2946-1-8 , pages =

    work page doi:10.1186/1758-2946-1-8

  41. [42]

    Robert and Abou-Shehada, Sarah and Dunn, Peter J

    Prat, Denis and Wells, Andy and Hayler, John and Sneddon, Helen and McElroy, C. Robert and Abou-Shehada, Sarah and Dunn, Peter J. CHEM21 selection guide of classical- and less classical-solvents. Green Chem. 2016. doi:10.1039/C5GC01008J

    work page doi:10.1039/c5gc01008j 2016

  42. [43]

    ACS Sustainable Chemistry & Engineering , volume =

    Bystrzanowska, Marta and Petkov, Petko and Tobiszewski, Marek , title =. ACS Sustainable Chemistry & Engineering , volume =. 2019 , doi =

    work page 2019

  43. [44]

    A framework for evaluating the chemical knowledge and reasoning abilities of large language models against the expertise of chemists , journal=

    Mirza, Adrian and Alampara, Nawaf and Kunchapu, Sreekanth and R. A framework for evaluating the chemical knowledge and reasoning abilities of large language models against the expertise of chemists , journal=. 2025 , day=

    work page 2025

  44. [45]

    1984 , doi =

    On a multicriteria shortest path problem , journal =. 1984 , doi =

    work page 1984

  45. [46]

    Multiobjective A* search with consistent heuristics , volume =

    Mandow, Lawrence and. Multiobjective A* search with consistent heuristics , volume =. J. doi:10.1145/1754399.1754400 , pages =

    work page doi:10.1145/1754399.1754400

  46. [47]

    and White, Chelsea C

    Stewart, Bradley S. and White, Chelsea C. , title =. 1991 , volume =. doi:10.1145/115234.115368 , journal =

    work page doi:10.1145/115234.115368 1991

  47. [48]

    Proceedings of the 14th Annual Conference on Genetic and Evolutionary Computation , pages =

    Brockhoff, Dimo and Wagner, Tobias and Trautmann, Heike , title =. Proceedings of the 14th Annual Conference on Genetic and Evolutionary Computation , pages =. 2012 , doi =

    work page 2012

  48. [49]

    A simple similarity metric for comparing synthetic routes

    Genheden, Samuel and Shields, Jason D. A simple similarity metric for comparing synthetic routes. Digital Discovery. 2025. doi:10.1039/D4DD00292J

    work page doi:10.1039/d4dd00292j 2025