PaperlayerView exact revision on arXiv

Source-linked AI summary

Beyond Drug Discovery: The Nanotechnology Molecular Optimization (NMO) Benchmark

Matthias Blaschke, Daniel Kienzle, Zsuzsanna Koczor-Benda, Julian Lorenz, Rainer Lienhart, Fabian Pauly

arXiv:2606.30170v1cs.LGcond-mat.mes-hallcs.AIcs.CEphysics.chem-ph

TL;DR

Existing molecular-design benchmarks and pharmaceutical pretraining favor proxy-task performance over transfer to structurally distinct nanotechnology problems. The paper introduces the quantum-simulation-based NMO Benchmark and a constraint-aware baseline, finding that state-of-the-art models struggle while a simple genetic algorithm remains competitive and the baseline surpasses prior physical-property results.

  • Problem

    Molecular-design research lacks benchmarks and models that evaluate generalization to scientifically grounded nanotechnology objectives beyond pharmaceutical proxy tasks.

  • Method

    The paper introduces NMO, using quantum simulations and strict protocols, and develops a constraint-aware baseline with a novel structural representation and domain-agnostic pretraining.

  • Results

    State-of-the-art generative models fail or struggle across NMO tasks, while a simple genetic algorithm remains competitive and the baseline surpasses prior results across all three physical-property oracles.

  • Takeaways & Limitations

    NMO provides a scientifically grounded testbed showing that simple methods can generalize where sophisticated pharmaceutical-oriented models do not.

  • Takeaways & Limitations

    xTB calculations trade accuracy for tractability, and the MO task may overestimate upconversion intensities in its high-fitness tail.

Abstract

from arXiv · show

Generative molecular design is shaped by simple proxy benchmarks for drug-like properties and models pretrained on large pharmaceutical datasets. This combination yields strong benchmark metrics but limits transferability to domains structurally distinct from drug discovery. To overcome this limitation and drive discovery toward real, scientifically grounded targets, we introduce the Nanotechnology Molecular Optimization (NMO) Benchmark, which bridges machine learning (ML) and quantum materials science. NMO acts simultaneously as a rigorous testbed for the ML community and a discovery engine for nanotechnology research. The suite replaces proxy oracles with quantum simulations and introduces strict protocols that prioritize scientific utility over leaderboard-oriented overfitting. The physics-based NMO tasks impose hard structural constraints and rugged fitness landscapes, posing fundamentally new requirements on generative models. Notably, advanced molecular optimization methods underperform much simpler approaches on the NMO tasks. We develop a new baseline method identifying the critical components to solve the NMO tasks, including a novel representation for modeling structural constraints and a domain-agnostic pretraining strategy to eliminate pharmaceutical dataset bias. Our results surpass state-of-the-art physical properties and reveal previously unknown structural motifs, offering new insights for the nanotechnology community and demonstrating that ML can drive genuine scientific discovery.

1 Introduction

Generative molecular design has achieved strong benchmark performance but remains concentrated in pharmaceutical applications and proxy objectives that inadequately represent real-world complexity. The NMO Benchmark addresses this gap by evaluating molecular design for nanotechnology applications involving quantum physics, structural constraints, and strict anti-overfitting protocols.

  • Motivation: Generative molecular design is increasingly specialized for pharmaceutical applications and proxy oracles that do not capture complex real-world objectives.The passage notes that seemingly generalist models are implicitly specialized and that recent methods optimize proxy tasks rather than structurally distinct discovery problems.
  • Motivation: Nanotechnology offers a bottom-up setting for tailoring individual molecules into functional devices, including Self-Assembled Monolayers and Metal-Organic Frameworks.The passage identifies these molecular systems as prominent nanotechnology examples and notes their recognition with the 2025 Nobel Prize in Chemistry.
  • NMO Benchmark Suite: The NMO Benchmark evaluates molecular design for quantum-physics applications spanning phonon transport, thermoelectric efficiency, and molecular optomechanics.Its molecular systems include single-molecule junctions for thermal or thermoelectric transport and surface-anchored molecules for THz detection via Raman scattering.
  • NMO Benchmark Suite: NMO enforces strict evaluation protocols against task-specific overfitting and incorporates hard structural constraints typical of real-world nanotechnology.The suite is designed to remain accessible and reproducible for machine-learning researchers without a physics background.

2 Related Work

Generative molecular design is dominated by pharmaceutical benchmarks that optimize drug-like properties through simple proxy oracles. Existing molecular representations and pharmaceutical-data-derived vocabularies can violate structural validity, introduce drug-like bias, and lack real-lab accessibility.

  • Methods in Generative Molecular Design: Pharmaceutical benchmarks such as PMO dominate molecular optimization by evaluating drug-like properties with simple proxy oracles.Examples include similarity to arbitrary substances.
  • Methods in Generative Molecular Design: SMILES permits syntactically correct strings that violate chemical rules.
  • Methods in Generative Molecular Design: SAFE uses fragments but does not guarantee valid combinations, while SELFIES and Group SELFIES rely on post-hoc parsers that can cause arbitrary structural truncations.
  • Methods in Generative Molecular Design: Models such as GenMol mine vocabularies from massive pharmaceutical datasets using BRICS heuristics, introducing implicit drug-like bias without guaranteeing real-lab accessibility.

3 The NMO Benchmark Suite

The NMO Benchmark Suite standardizes generative molecular design evaluation across three quantum-materials applications under a strict 10,000-oracle-evaluation budget. It combines physics-based fitness and synthesis constraints with specialized structural representations and anti-exploitation protocols.

  • Benchmark scope: NMO evaluates generative molecular design across Phonon Transport (PH), Thermoelectrics (TE), and Molecular Optomechanics (MO).The suite is designed to be accessible to the ML community without requiring prior quantum-physics expertise.
  • Task formulation: Each task maximizes a physics-based fitness function combining the task property, synthetic-accessibility and rotatable-bond penalties, and hard physical constraints.The benchmark limits optimization to 10 000 oracle evaluations because quantum simulations and experimental validations are costly.
  • Evaluation protocol: Success is measured by Top-10 AUC, Mean Top-10 Fitness, and Mean Top-10 SA under a strict budget of N = 10000 fitness evaluations.These metrics assess sample efficiency, candidate quality, and the synthesizability of the best candidates.
  • Molecular representation: Graph Group SELFIES (GGS) provides an optional interface for representing electrode-attachment anchors while avoiding structural truncation caused by unavailable attachment points.NMO accepts standard SMILES with explicit atomic indices, but standard strings lack native topological support for anchors.
  • Evaluation protocol: The evaluation protocol forbids task-specific datasets and hyperparameter tuning, requiring one configuration across all three tasks to prevent benchmark exploitation.The suite also provides a SMILES-translated synthetic pretraining dataset and permits bounded molecule lengths for PH to avoid a known oracle degeneracy.

4 Methodology

The methodology evaluates five popular molecular-optimization methods spanning dominant paradigms, including sequence-model reinforcement learning and training-free genetic search. It also reports literature-method and ablation results on NMO using SA scores, with values above 4.5 highlighted.

  • Literature Methods: Five popular molecular-optimization methods are evaluated to cover the dominant paradigms.The suite includes REINVENT, MolGA, and GenMol among the evaluated methods.
  • Literature Methods: REINVENT pre-trains a molecular sequence model and fine-tunes it with reinforcement learning using oracle scores as rewards.It is built around the SMILES encoding, which can be adapted to novel molecular encodings.
  • Literature Methods: MolGA is a training-free genetic algorithm that generates molecular offspring through crossover and mutation.Like REINVENT, it is built around the SMILES encoding but can be adapted to novel molecular encodings.
  • Evaluation: Table 1 evaluates literature methods and an ablation study on the NMO benchmark using SA scores, with scores exceeding 4.5 highlighted in red.Zero values in the table are attributed to rounding.

5 Experiments

Experiments show that pharmaceutical-pretrained literature methods fail on all three NMO oracles, whereas the baseline finds candidates surpassing prior results. The evaluation also identifies practical limitations involving oracle reliability, fragment-library bias, and computational cost.

  • Method comparison: f-RAG and GenMol fail across all three oracles with near-zero performance, indicating poor adaptation from pharmaceutical pretraining to physical objectives.Both methods use SMILES and massive pharmaceutical pretraining, with default vocabularies and implicit first/last non-hydrogen-atom anchors.
  • Ablations: Switching from ZINC pretraining to GGS brings SA below threshold, while auxiliary descriptors measurably degrade MO performance.The GGS change costs some AUC and fitness, and the auxiliary objective appears to conflict with MO optimization.
  • Baseline results: The baseline identifies top-performing candidates across all three oracles, surpassing previous literature results in each respective field.Selected examples are presented in Figure 3, with detailed physical analysis provided separately.
  • Limitations: xTB calculations may overestimate MO upconversion intensities for P > 15, although explicit DFT checks still indicate strong performance at smaller P values.The xTB approach trades accuracy against tractability, so high-fitness-tail candidates require careful interpretation.
  • Limitations: The fragment library introduces controllable bias, while high computational cost restricts evaluation to five seeds and reduces statistical power.The five-seed evaluation matches or exceeds standard practice for cheaper benchmarks.

6 Conclusion

The NMO Benchmark introduces a nanotechnology-focused testbed for generative molecular design that replaces proxy oracles with quantum simulations and rewards generalist methods over per-task overfitting. Its scientifically grounded tasks remain accessible to ML researchers while imposing rugged fitness landscapes and hard structural constraints.

  • NMO introduces a benchmark for generative molecular design in nanotechnology, grounding its tasks in relevant scientific problems.
  • Quantum simulations replace proxy oracles, while a strict protocol rewards generalist methods over per-task overfitting.
  • The benchmark remains accessible to ML researchers without a physics background while posing rugged fitness landscapes and hard structural constraints.

Technical appendices and supplementary material … Details on underlying theory

The supplementary material develops the physical theory and implementation of the three NMO tasks, validates the implemented oracles, contextualizes results, and discusses limitations. For heat transport, it formulates phonon conductance through transmission, molecular Green’s functions, electrode coupling, and a Debye-model surface Green’s function.

  • Technical appendices and supplementary material: The appendices explain the physical background, practical implementation, oracle validations, literature context, and limitations for the three NMO benchmark tasks.This material supports both methodological interpretation and assessment of the implemented physics-based oracles.
  • A NMO Benchmark: The NMO task details connect benchmark objectives to physical theory and implementation choices for scientifically grounded molecular optimization.The section explicitly frames the task descriptions around theory, implementation, validation, context, and limitations.
  • A.1.1 Heat Transport at the Single Molecule Level: The heat-transport task targets molecular junctions that efficiently conduct or insulate heat for thermal management and heat-based devices.Thermal conductance is measured in Watt per Kelvin (W/K), representing transmitted power per temperature difference between electrodes.
  • A.1 NMO Task Details: Phonon-mediated transport is modeled with the Landauer-Büttiker formalism, balancing numerical accuracy against computational time through adapted numerical schemes.The implementation adapts a formalism used in prior work and recognizes trade-offs between physical rigour and computational demand.
  • Details on underlying theory: The thermal conductance κph integrates the energy-dependent phonon transmission τph(E), weighted by the temperature derivative of the Bose-Einstein distribution.τph(E) describes phonon transmission probability, while ∂n(E,T)/∂T reflects the quantum nature of phonon transport.
  • Details on underlying theory: Molecular structural changes mainly affect transmission locally through quantum interference, making large conductance variations difficult because conductance integrates transmission.The difficulty arises from the integral dependence of thermal conductance on transmission.
  • Details on underlying theory: Phonon transmission is computed from molecular Green’s functions, linewidth broadening matrices, the mass-weighted Hessian, and electrode self-energies.The Green’s function captures molecular scattering, while linewidth matrices describe electrode coupling.
  • Details on underlying theory: Electrode self-energies depend on a Debye-model surface Green’s function, avoiding explicit electrode regions and significantly reducing simulation computational demands.The Debye model is presented as a minimalistic model for phonons in metals.

Details on practical implementation and validation · A.1.2 Thermoelectric Efficiency at the Molecular Scale · Details on underlying theory

The NMO workflow combines gold-thiol molecular encoding, geometry optimization, automated code, and validation against substituted molecular junctions. Its thermoelectric framework models coupled electronic and thermal transport using transmission functions, realistic electrode structure, and molecular-orbital effects.

  • Details on practical implementation and validation: Gold-thiol anchors attach to the first and last non-hydrogen SMILES atoms, followed by xtb geometry optimization toward a minimum-energy configuration.The encoding defines anchor positions operationally for SMILES strings and uses experimentally and theoretically established gold-thiol anchors.
  • Details on practical implementation and validation: Complete program code automates the workflow, requiring only a molecule’s GGS or SMILES string and enabling non-experts to use the benchmark.
  • Details on practical implementation and validation: Validation uses pBDA junctions and halogen substitutions—fluorine, chlorine, bromine, and iodine—to examine interference effects in phonon transmission.The substitutions replace one benzene-ring hydrogen and increase in mass across the halogen series.
  • Details on practical implementation and validation: Highly relevant NMO molecules have thermal conductance below 0.25 pW/K and SA score below 4.5, emphasizing synthetic accessibility.Literature candidates optimized for low thermal conductance reach κph = 0.1 −0.4 pW/K and often have SA scores around 4 or above 4.5.
  • A.1.2 Thermoelectric Efficiency at the Molecular Scale: High thermoelectric efficiency requires high electrical conductance, high Seebeck coefficient, and low thermal conductance, despite intrinsic coupling between transport coefficients.Single-molecule junctions may mitigate these correlations through discrete molecular orbitals acting as sharp energy filters.
  • Details on underlying theory: Electronic transport follows the Landauer–Büttiker formalism, with conductance determined by electronic transmission near the electrode Fermi energy.The derivative of the Fermi–Dirac distribution is sharply peaked near µ ≈EF at low temperature, restricting the relevant energy window.
  • Details on underlying theory: The Seebeck coefficient measures thermoelectric voltage per temperature difference and typically reaches several tens of µV/K in single-molecule junctions.It is measured in Volts per Kelvin (V/K).
  • Details on underlying theory: The method computes molecular Hamiltonians and overlap matrices with xtb, incorporates explicit electrode structure for level alignment, and reuses surface Green’s functions for high-throughput transport calculations.Electrode structure establishes a well-defined Fermi energy, while recursive Green’s-function calculations are optimized for GPUs and CPUs.

Limitations

The approach faces limitations in accurately predicting the Seebeck coefficient because it is highly sensitive to molecular energy-level alignment at the Fermi energy. Nevertheless, the method remains predictive and standard for electronic transport modeling in single-molecule junctions.

  • Limitations: The Seebeck coefficient is particularly challenging to predict accurately because it depends on the transmission-function slope at the Fermi energy.Its sensitivity arises from the precise alignment of molecular energy levels relative to EF.
  • Limitations: Gating could fine-tune the molecular energy-level alignment in experimental settings.
  • Limitations: Explicit electrode inclusion establishes a solid foundation for the modeling approach.
  • Limitations: Despite these shortcomings, the approach has demonstrated considerable predictive power and remains standard for modeling electronic transport in single-molecule junctions.

Details on practical implementation and validation … A.3.4 Selection of Building Blocks

The appendix details automated, validated quantum-property workflows, task-specific thresholds and filters, structured molecular representations and genetic operations, and domain-informed building-block selection. These design choices constrain chemically valid exploration while retaining search spaces large enough to support physics-based discovery.

  • Details on practical implementation and validation: The implementation provides reproducible code and scripts, automates workflows from GGS or SMILES strings, and attaches gold clusters for electronic-transport calculations.The workflow also calculates electronic transport quantities after constructing the electrode–molecule junction.
  • Details on practical implementation and validation: Benchmark validation reproduces characteristic pOPE/mOPE transmission features, including the para-to-meta destructive-interference dip and the literature ordering of side-group-dependent features.The transmission near the Fermi energy determines conductance, while side groups shift interference-feature energies.
  • Details on practical implementation and validation: The HBT-OPE3-A validation gives measured ZT ≈1.3×10−5 and method ZT = 1.52 × 10−3, while quantitative agreement with experiments remains challenging.The reported measurement used G = 1.82×10−4 G0, S = −8.7 µV/K, and κ = 24 pW/K; the method used G = 2.3 × 10−3 G0.
  • Details on underlying theory: For the optomechanics task, the workflow computes P and I_c^m using PTB and GFN2-xTB, models molecular geometry through endpoint alignment, and automates relaxation and alignment from GGS or SMILES.The molecular P property aggregates modes in the 30-1000 cm−1 THz-MIR range, and the source endpoint defines gold–thiol attachment while the endpoint axis is aligned with z.
  • A.1.4 Thresholds for Relevant Molecules: Promising candidates are defined by task-specific physical thresholds combined with SA < 4.5: k_ph < 0.25 pW/K for PH, ZT > 3 for TE, and P > 7.88 for MO.The analysis scripts automatically check these thresholds for generated molecules.
  • A.2 Molecular Filters: SMARTS preprocessing filters exclude reactive or unstable substructures before oracle calls, while a HOMO-LUMO-gap criterion supplies an additional hard constraint and sampling distributions improve candidate discovery.The SMARTS filters do not contribute to the computational budget, but they cannot cover every undesirable motif.
  • A.3 Encoding Details: GGS represents molecules as directed acyclic graphs with SELFIES-compatible textual notation; its current implementation has maximum branching depth one and omits chirality and stereochemistry tokens.Genetic search uses rank-based parent selection, crossover, and mutation operations that alter fragments, bonds, anchors, branches, termini, and chain length.
  • A.3.4 Selection of Building Blocks: Building blocks were expanded to cover PMO oracle targets and molecular-phononics motifs, while sulfur-containing fragments were excluded from physics-based backbones because sulfur is reserved for gold-binding anchor groups.The resulting search space is estimated to be ≫1030, and domain-informed selection focuses exploration despite limiting generated molecular types.

A.4 Creation of Synthetic Datasets … B.6 Full Pipeline

The paper presents a procedurally generated, bias-controllable molecular-design pipeline for NMO, combining constrained GGS construction, optional chemical filtering, descriptor-aware pretraining, and adaptive optimization. Its key findings emphasize anchor-aware modeling, task-robust design, rugged-landscape stabilization, and practical safeguards for physically grounded discovery.

  • A.4 Creation of Synthetic Datasets: Synthetic datasets use stochastic directed-acyclic-graph assembly under fragment-count and [pop]-operation limits, with actions sampled from an attachment-point grammar.The grammar begins from a seed fragment and iteratively adds fragments and edges.
  • A.4 Creation of Synthetic Datasets: The pipeline supports explicit fragment-vocabulary control, optional chemical-stability filtering, and dual GGS/SMILES output for data-scarce physical domains.Custom vocabularies can match laboratory reagents and synthesis capabilities, while GGS guarantees syntactical validity and filters improve chemical plausibility.
  • A.5 Code and Data Availability: The codebase is openly available under an MPL-2.0 license and includes the benchmark, GGS encoding, baseline models, training scripts, evaluation scripts, documentation, and READMEs.The repository is designed for community use and extension.
  • A.6 Solving NMO: Key Findings: Anchor positions are decisive because NMO properties depend on the full molecule and its electrode-binding anchors, motivating anchor-aware GGS or SMILES representations.The authors state that anchors should be part of the learning process rather than auxiliary metadata.
  • A.6 Solving NMO: Key Findings: Chemical-space design affects optimization: MO benefits from pharmaceutical bias, whereas TE and PH do not, so no single prior suits every task.The benchmark provides curated building blocks and reports the effect of removing critical fragments such as acetylene.
  • A.6 Solving NMO: Key Findings: The benchmark favors robust single configurations across three physically distinct tasks, while dynamic stabilization mechanisms address rugged landscapes without KL anchoring.Rugged landscapes can cause catastrophic forgetting or mode collapse, especially for TE; the reported mechanisms are DCD and DEX.
  • A.7 Impact Statement: The framework broadens molecular engineering beyond dataset bias through custom fragment vocabularies adaptable to nanotechnology, materials science, physics, chemistry, life sciences, and laboratories.The impact statement frames this flexibility as useful across scientific domains and experimental constraints.

B.7 Methodological Details: Action Space · B.8 Extended Evaluation · B.8.1 Pretraining

The method constructs molecules through indexed actions that begin at a start state and terminate with an [end] token, using encoding-specific action spaces for SMILES and GGS. Extended evaluation shows that GGS simplifies pretraining by representing only valid molecules, while descriptor prediction rapidly improves and then saturates during pretraining.

  • B.7 Methodological Details: Action Space: Molecular construction iteratively samples indexed actions from an encoding-specific action space, beginning at a start state and terminating with [end].
  • B.7 Methodological Details: Action Space: SMILES actions are drawn from a fixed vocabulary of atomic, bond, and ring-closure tokens that defines the chemical space.
  • B.7 Methodological Details: Action Space: The primary SMILES vocabulary contains 53 distinct tokens, with additional dataset-specific tokens added when needed; pretraining must learn chemically valid SMILES construction.
  • B.7 Methodological Details: Action Space: GGS actions comprise fragments, coupling-in and coupling-out options, pop, and end tokens, with coupling options determined by fragment-library attachment points.
  • B.8 Extended Evaluation: Extended evaluation examines synthetic-dataset pretraining, the interaction between agent sampling and genetic search, and stability mechanisms.
  • B.8.1 Pretraining: GGS pretraining is considerably easier than SMILES pretraining because GGS represents only valid molecules, leaving the agent to learn valid GGS-string syntax rather than underlying chemistry.
  • B.8.1 Pretraining: The original Genetic GFN’s RNN with SMILES encoding has the lowest valid rate after 5 epochs, while switching to the random dataset increases learning speed and valid rate.
  • B.8.1 Pretraining: A non-interfering regression head predicts K = 17 molecular descriptors; normalized descriptor loss drops rapidly during the first epoch and then saturates with a slight downward trend.

B.8.2 Sampling and Genetic Algorithm … B.9.1 Genetic GFN Framework Components

Across the NMO oracles, hybrid agent sampling with genetic refinement is essential, while graph-based GGS operators outperform SMILES-based genetic search by preserving valid molecular structures. Ablations further show that GGS resolves the topology mismatch for thermoelectric optimization, whereas adaptive mechanisms improve training stability by reducing invalid sampling.

  • B.8.2 Sampling and Genetic Algorithm: For thermoelectric optimization, only a fraction of 10000 tested molecules satisfy the hard constraints, and SMILES-based genetic search fails to identify top-performing candidates.The proposed method produces significantly more candidates with positive fitness than the original Genetic GFN.
  • B.8.2 Sampling and Genetic Algorithm: For phonon transport, few candidates satisfy the hard constraints, SMILES-based genetic search is ineffective, and the proposed method observes numerous candidates with fitness > 0.Initially, high-performing candidates are discovered through both agent sampling and genetic search.
  • B.8.2 Sampling and Genetic Algorithm: For the optomechanical oracle, hard constraints are relatively easy to satisfy, making string-based genetic search more effective; both agent sampling and genetic search remain important.The larger number of valid molecules fills the replay buffer and improves string-based search.
  • B.8.2 Sampling and Genetic Algorithm: Hybrid agent sampling and genetic refinement are key across all oracles, and the proposed genetic search performs significantly better than SMILES-based search.The graph-based approach and GGS encoding make genetic operators yield valid molecules, although hard constraints can still produce zero-fitness offspring.
  • B.8.3 Stability via Adaptive Hyperparameters: Switching to Dynamic Cooldown and Dynamic Exploration halves the invalid rate for the thermoelectric oracle.The training analysis tracks invalid and duplicate rates with and without these adaptive mechanisms.
  • B.9 Ablation Studies: The ablation study evaluates each genetic GFN extension using performance, molecule-discovery, and invalid-rate metrics, reflecting stability challenges introduced by its sampling step.Invalid rate is defined as the percentage of invalid samples in the final step.
  • B.9.1 Genetic GFN Framework Components: SMILES variants fail to find high-performing thermoelectric candidates because they cannot natively model two-sided gold binding, whereas GGS resolves this topology mismatch and substantially increases AUC and fitness.GGS also enables robust RI discovery; the transformer architecture leaves thermoelectric performance unchanged within noise, while invalid rates remain high.

B.9.2 Fragment Library · B.10 Analysis of the Best Performing Candidates

The fragment library strongly influences NMO performance: acetylene is critical across tasks, especially phononic-transport tasks, although important motifs can still emerge without it. The best-performing candidates are further analyzed through their physical properties and theoretical explanations, with all RI-qualified baseline candidates provided.

  • B.9.2 Fragment Library: Acetylene is a critical motif across all tasks, and its inclusion reflects domain knowledge about beneficial properties in molecular junctions.The fragment library was selected using domain knowledge, and prior work reports beneficial acetylene properties in molecular junctions.
  • B.9.2 Fragment Library: Removing acetylene from the TE library lowers AUC and mean fitness but improves top-candidate SA, leaves invalid rate unchanged, and preserves candidates surpassing the RI threshold.TE benefits from low phononic transport, for which acetylene is expected to help.
  • B.9.2 Fragment Library: Removing acetylene most strongly harms PH, lowering AUC and mean fitness, slightly reducing invalid rate, preserving SA, and eliminating candidates that pass the RI threshold.Its absence makes discovering high-performing candidates more difficult because acetylene supports low phononic transport.
  • B.9.2 Fragment Library: For MO, acetylene removal lowers AUC and mean fitness but improves top-candidate SA, leaves invalid rate unchanged, and preserves conjugated-chain and aromatic-ring motifs.The resulting molecules retain the same broad structural motifs without acetylene itself.
  • B.9.2 Fragment Library: The fragment library introduces inherent bias, motivating future open-vocabulary methods that discover important motifs such as acetylene without receiving them as building blocks.The benchmark components are discussed to enable methods that perform well while reducing dependence on a fixed fragment vocabulary.
  • B.10 Analysis of the Best Performing Candidates: The paper analyzes the best-performing candidates in depth by examining their physical properties and theoretically grounding the reasons for their outstanding performance.This section focuses on candidates identified by the proposed method.
  • B.10 Analysis of the Best Performing Candidates: All candidates proposed by the baseline that exceed the RI threshold are provided in addition to the selected candidates analyzed in detail.The complete RI-qualified candidate set is included alongside the presented selection.

B.10.1 Thermoelectric oracle … B.11.1 Compute Resources

The paper analyzes top candidates for thermoelectric, phonon-transport, and molecular-optomechanics optimization, then details the training setup and computational requirements. The candidates achieve strong simulated properties, while validation also exposes limitations of proxy predictions and the computational bottleneck of quantum-chemistry evaluations.

  • B.10.1 Thermoelectric oracle: Suppressed phononic transport arises from acetylene groups near anchoring sites and destructive quantum interference induced by amino or pyrimidine side groups.These features are identified as critical for the candidate’s thermoelectric performance.
  • B.10.1 Thermoelectric oracle: The algorithm discovers combinations of literature-studied molecular motifs without explicit prior knowledge and within a very limited number of oracle calls.This finding concerns the thermoelectric multi-objective optimization problem.
  • B.10.1 Thermoelectric oracle: The thermoelectric candidate is promising despite experimental ZT values likely being lower, exceeding the comparable theoretical maximum of ZT = 2.4 while retaining reasonable synthetic accessibility.The candidate uses standard chemical building blocks and combines robust transport features.
  • B.10.2 Phonon Transport Oracle: 0.0990 pW/K is the phonon candidate’s calculated room-temperature thermal conductance, reported as κph = 0.0990 pW/K.The candidate is evaluated through its phonon transmission and thermal conductance.
  • B.10.3 Molecular Optomechanics Oracle: The selected optomechanics candidate has P values of 9.91 by PTB and 8.31 by DFT, an SA score of 4.35, and a shape supporting SAM and nanocavity formation.Selection balanced target-property values, synthetic accessibility, and molecular shape rather than fitness alone.
  • B.10.3 Molecular Optomechanics Oracle: DFT validation of 180 top-performing molecules found 78 with P values above the prior best candidate’s P = 7.88, while PTB severely overestimated P for some molecules.The validation assesses the reliability of the PTB method for molecular optomechanics.

B.12 Experiments on the PMO Benchmark

The baseline method is also evaluated on the 23-task PMO benchmark for drug-like property optimization. Results examine literature comparisons, oracle-query limitations, and the effects of ZINC pretraining and model design choices.

  • Benchmark evaluation: The baseline method is applied to the Practical Molecular Optimization benchmark, which contains 23 tasks evaluating generative models on drug-like properties.PMO complements the NMO-focused experiments with an established drug-discovery benchmark.
  • Benchmark protocol: f-RAG and GenMol use task-specific vocabularies and hyperparameters built by heavily querying the oracle, undermining the 10,000-call limit and initializing optimization near optimal solutions.This setup prevents meaningful assessment of actual optimization performance for these methods.
  • Pretraining ablation: Genetic GFN performance drops significantly without ZINC-based pretraining, while original-publication values differ slightly, likely because of random seeds.The comparison includes the original Genetic GFN pretraining dataset and ablations from Table 1.
  • Comparative performance: Across ablation variants, the Genetic GFN family performs on par with molGA and REINVENT on PMO, remaining competitive across encodings and architectural choices.The comparison supports the competitiveness of the underlying framework beyond a single representation or architecture.
Loading 2606.30170v1…