Retrosynthesis planners — search programs that work backwards from a target molecule to reactions and buyable starting materials — have a blind spot that is not algorithmic but physical. They are trained on patent data, and patent data comes overwhelmingly from batch chemistry: flasks where every reagent, intermediate, and byproduct coexists for minutes to hours. Under that regime, any synthetic route that passes through a short-lived intermediate is genuinely a bad idea, so such routes never enter the data, and planners never propose them. Yet in a continuous-flow reactor, reaction time is mapped onto position: the time a molecule spends in a reaction zone is the zone's volume divided by the pump rate, adjustable from hours down to fractions of a second. With that single knob, chemists have already demonstrated reactions that are impossible in batch — an unstable intermediate is generated in one zone and consumed in the next before it can degrade (Multistep Synthesis of Paracetamol in Continuous Flow); even a carbocation and a carbanion, which destroy each other on contact in a flask, can be made in separate streams and coupled in milliseconds (Convergent approach for direct cross-coupling enabled by flash irreversible generation of cationic and anionic species).
The core claim of this proposal is that mode collapse in retrosynthesis — the search concentrating on one narrow family of solutions — is, in this domain, partly a data-and-physics problem whose missing modes are exactly these transient-intermediate routes. The fix is not another exploration heuristic on top of a batch-biased search space, but a change to what "feasible" means: chemical plausibility should be a function of residence time $\tau$, not a binary property of the reaction graph. We propose to build a $\tau$-conditioned plausibility oracle, plug it into diversity-aware route search, and show that the search then recovers route families that every existing planner prunes by construction.
Background: Why Current Planners Cannot Find These Routes
Three lines of work define the frontier, and each stops short of the intersection. Diverse tree search is represented by $DFPN^*$, which prevents mode collapse in multistep retrosynthesis by penalizing proof and disproof numbers along already-found routes and disproving their terminal nodes, forcing the tree into novel disconnections (Completeness and Diversity in Depth-First Proof-Number Search with Applications to Retrosynthesis). Diverse generative models are represented by RetroGFN, which samples synthetic trajectories proportional to a chemical reward rather than maximizing likelihood, and RSF, which steers discrete flow matching with sequential Monte Carlo resampling (Diverse and Feasible Retrosynthesis using GFlowNets; RETRO SYNFLOW: Discrete Flow-Matching for Accurate and Diverse Single-Step Retrosynthesis). Neither line models process chemistry at all — reactions are static graph edits, feasibility is a learned score over patent statistics. Meanwhile, the flow-chemistry literature demonstrates exquisite temporal control but optimizes one hand-designed route at a time, with manual parameter sweeps (Multistep Synthesis of Paracetamol in Continuous Flow).
The sharpest evidence that the two worlds conflict is a direct contradiction between two cited papers. ChemCensor, the plausibility metric introduced with the CREED dataset, explicitly rejects proposed reactions in which reactive intermediates coexist with incompatible functional groups (When Single Answer Is Not Enough: Rethinking Single-Step Retrosynthesis Benchmarks for LLMs). The flash cross-coupling paper proves that exactly such coexistence — cationic and anionic species meeting in a micromixer — gives yields up to 96% when the coexistence time is under ten seconds (Convergent approach for direct cross-coupling enabled by flash irreversible generation of cationic and anionic species). Incompatibility is fatal only above a timescale, and that timescale is an engineering variable. Note that "flow" in RSF (RETRO SYNFLOW) is purely mathematical — flow matching in generative modeling — with no connection to fluid flow; we use "flow" here only in the physical sense.
Research Program
The physical lever is elementary: in a flow reactor, residence time $t$ in a zone of volume $V$ at volumetric flow rate $Q$ is $t = V/Q$, so zone lifespans are set with millisecond-to-second precision. The program has three parts.
First, a $\tau$-conditioned plausibility oracle. We formalize the physics as a feasibility window: a reaction step that forms transient intermediate $I$ is $\tau$-feasible if the mixing time needed to form and consume it, $t_{\text{mix}}$, is shorter than the residence time, which in turn is shorter than the time for $I$ to degrade along its dominant side pathway, $t_{\text{decay}}(I)$:
$$t_{\text{mix}}(r) \;<\; \tau \;<\; t_{\text{decay}}(I_r).$$
The oracle predicts $t_{\text{decay}}$ from molecular structure using reactive-species classifiers (carbocations, carbanions, $N$-arylhydroxylamines and related species) trained on the 22.7 million reactions of CREED (When Single Answer Is Not Enough), with a required consistency property: as $\tau \to$ hours it must reproduce batch plausibility (agreement with ChemCensor labels), and at $\tau < 1$ s it must rank the literature flow reactions correctly. Calibration uses the published flow case studies — the paracetamol platform, where $N$-arylhydroxylamine selectivity falls from above 98% at 12 seconds to 79.1% at 60 seconds (Multistep Synthesis of Paracetamol in Continuous Flow), and the convergent coupling set, with yields of up to 83%, 72%, 96%, and 92% at residence times from 0.36 to 21 seconds (Convergent approach for direct cross-coupling). We are explicit about a data limitation: TotSyn, the largest curated total-synthesis dataset (105,002 reactions, 30,703 templates), carries no residence-time labels (TotSyn: A Total Synthesis Reaction Dataset for Machine Learning in Organic Chemistry), so the oracle must be seeded from these small published flow sets plus physical priors, and its calibration is a first-class research question, not an afterthought.
Second, flow-augmented diverse route search. We extend the $DFPN^*$ search backbone with a library of flow-enabled operators — convergent coupling of electrostatically incompatible partners, generation-and-immediate-consumption cascades like the hydrogenation-to-Bamberger-rearrangement sequence in the paracetamol synthesis — each gated by the oracle's $\tau$-feasibility check. Crucially, the diversity penalty is applied over (route, $\tau$-regime) pairs: the search score for candidate $x$ becomes the base feasibility score minus a coefficient times its similarity to an archive $\mathcal{M}$ of already-accepted (route, regime) solutions, in the spirit of DARWIN's hard Tanimoto bound and SGRPO's leave-one-out group credit (DARWIN: A Framework for Target Specific Diversity Constrained Natural Product Like Molecule Generation; Pushing Biomolecular Utility-Diversity Frontiers with Supergroup Relative Policy Optimization). This is needed because the expanded space is even more prone to collapse than the batch space: once the search finds one well-characterized fast regime, unpenalized search will saturate it.
Third, and secondary, we apply the same anti-collapse machinery to the reactor side: multi-zone flow parameter search (flow rates, volumes, temperatures per zone) where a naive optimizer collapses onto a single safe operating regime, and similarity-penalized search should instead surface distinct Pareto-optimal regimes — an ultra-fast high-temperature mode versus a moderate long-residence mode. The paracetamol platform's published sensitivity data (98% selectivity at 12 s versus 79.1% at 60 s) gives the ground-truth shape of this landscape (Multistep Synthesis of Paracetamol in Continuous Flow).
Proposed Experiments
The first experiment calibrates the oracle: on the convergent-coupling reaction set, it must recover the correct ordering of viable $\tau$-windows against the reported yields (rank correlation against the published 0.36–21 s windows and 72–96% yields), and on the paracetamol case it must predict that selectivity degrades from the 12-second to the 60-second regime (Multistep Synthesis of Paracetamol in Continuous Flow; Convergent approach for direct cross-coupling). The second experiment is the headline route-search comparison on the 60-molecule retrosynthesis benchmark (40 synthesizability molecules plus 20 Bayer drug-like molecules) used by $DFPN^*$, reporting solved molecules, unique intermediates per target, mean route length, and a new metric — the fraction of accepted routes that pass through a transient intermediate gated by a sub-second $\tau$-window, which we call flow-enabled route fraction. No reported result found on Lacuna for flow-enabled route fraction; the metric is introduced here, and for every existing planner it is zero by construction.
The third experiment is the honest control described under "Why not just…?" below: large candidate pools from off-the-shelf diverse generators, filtered post hoc by the oracle. The fourth experiment runs the multi-modal reactor parameter search on a simulated two-zone system calibrated to the paracetamol data, reporting the number of distinct feasible operating regimes and their spread in $(Q, V, T)$ space; no reported result found on Lacuna for automated multi-modal flow parameter search. The closing decisive check is a small wet-lab validation: two or three oracle-endorsed routes that ChemCensor rejects at batch timescales, executed in a flow reactor, with success defined as isolated product at yields comparable to the published flash-coupling envelope (tens of seconds, yields above roughly 70%).
Baselines
The do-nothing reference point is standard Monte Carlo tree search retrosynthesis planning, whose behavior on the exact 60-molecule benchmark is documented in the $DFPN^*$ paper: MCTS solved 38 of 60 molecules and generated an average of 101 unique intermediate molecules per target, with routes averaging 6.0 reactions (Completeness and Diversity in Depth-First Proof-Number Search with Applications to Retrosynthesis). It matters because it shows what generic exploration bonuses deliver — minor variations on one intermediate skeleton — and our method must clear it trivially or something is broken. The strongest existing competitor is $DFPN^*$ itself, the diversity-penalized proof-number search that solved 41 of 60 molecules, found 347 unique intermediates per target on average, and produced shorter routes (4.9 reactions on average) on the same benchmark with the same template library (Completeness and Diversity in Depth-First Proof-Number Search). We will re-run it ourselves under identical compute, template library, and tuning budget, so the only difference is the $\tau$-oracle gating and the flow-enabled operator library. Beating it means: at least as many molecules solved, at most its route length, more unique intermediates per target than its reported 347, and — the decisive part — a nonzero flow-enabled route fraction where its re-run has zero. For single-step generation, RetroGFN provides the diverse-generator reference on USPTO-50k, reporting 63.6% round-trip accuracy at $k=50$ versus approximately 54% for the next best model, at roughly 3.7 seconds per molecule (Diverse and Feasible Retrosynthesis using GFlowNets); RSF's steered variant reports 80.1% top-5 round-trip accuracy and 91.4% top-1 round-trip coverage on the same benchmark (RETRO SYNFLOW). These are used in the post-hoc filtering control, not as route-search competitors, since they are single-step models. Finally, two ablations isolate the proposal's ingredients: flow-aware greedy search (the $\tau$-oracle and operator library, but the diversity penalty removed) tests whether anti-collapse machinery is needed at all, and diverse search without the oracle (unmodified $DFPN^*$) tests whether the oracle is the active ingredient. If flow-aware greedy search matches the full method's flow-enabled route fraction, we will report that the mode-collapse story is unnecessary and the contribution reduces to conditioning feasibility on $\tau$ — a smaller but still useful claim.
Why not just…?
Why not just raise the sampling temperature or add an entropy bonus? In structured discrete domains this fails for a specific reason: random perturbations of valid molecules and reaction graphs mostly produce invalid or nonsensical candidates, so the extra samples are not new modes, just noise (AlphaSAGE: Structure-Aware Alpha Mining via GFlowNets for Robust Exploration), and unconstrained high-temperature generation is documented to produce the very functional-group incompatibilities that plausibility metrics reject (When Single Answer Is Not Enough). More fundamentally, temperature changes how a fixed prior is sampled, but the missing modes are absent from the prior itself and actively vetoed by the feasibility filter. The post-hoc control experiment measures this directly.
Why not just generate a huge pool and filter for flow compatibility afterwards? This is the most serious alternative, and we treat it as a genuine threat rather than arguing around it. The evidence against it is that collapse happens during generation, before any filter can act: standard policies collapse to one mode and then yield ten thousand near-identical samples (Pushing Biomolecular Utility-Diversity Frontiers with Supergroup Relative Policy Optimization), and patent-likelihood training prunes valid diverse routes that the data never contained (Diverse and Feasible Retrosynthesis using GFlowNets). But the counter-argument — that a $k=50$ pool from a good diverse model might already contain flow-viable transient-intermediate routes — is empirically checkable, and our third experiment does exactly that: pass RetroGFN and RSF pools through the oracle and count. If a meaningful number of flow-viable routes survive, we will say so plainly and the proposal's search component loses its motivation, leaving the oracle as the contribution.
Why not just use standard MCTS exploration, which already balances exploration and exploitation? The head-to-head evidence is unambiguous on the benchmark we adopt: MCTS returned 101 unique intermediates per target against 347 for diversity-penalized $DFPN^*$, solved fewer molecules (38 versus 41), and needed longer routes (6.0 versus 4.9 reactions) (Completeness and Diversity in Depth-First Proof-Number Search). MCTS mode-collapses onto minor variations of one intermediate skeleton; adding flow options to that search would multiply the space without fixing the collapse.
Why not just stabilize the intermediates with cryogenic batch chemistry instead of building flow modeling into the planner? The flow literature addresses this head on: the paracetamol intermediate cannot survive any one-pot batch arrangement because its formation needs neutral conditions while its consumption needs concentrated acid, and only physically separated flow zones resolve the conflict, giving 86% overall yield and 99.3% purity (Multistep Synthesis of Paracetamol in Continuous Flow); the cation-anion couplings failed in batch because dimerization outruns flask-scale mixing, while microflow mixers complete them in 0.36 to 21 seconds (Convergent approach for direct cross-coupling). Cryogenic batch cannot reach sub-second mass transfer, so a planner restricted to batch physics — even one with a cold-temperature knob — still cannot represent these route families.
Decisive Checks and Risks
The proposal lives or dies on two checks. The oracle calibration is the weakest link: if predicted $\tau$-windows do not rank correctly against the published flow reactions, nothing downstream is credible, and we would report the failure and stop. The post-hoc filtering control is the conceptual risk: it could show that existing diverse generators already propose flow-viable routes, in which case the search contribution collapses to an engineering detail. Both checks are scheduled first, before any wet-lab work. If both pass, the wet-lab validation of oracle-endorsed, batch-rejected routes — with the published flash-coupling yields of 72–96% at sub-ten-second residence times as the performance envelope (Convergent approach for direct cross-coupling) — is the final, decisive evidence that conditioning plausibility on residence time genuinely unlocks synthetic modes that a decade of batch-trained planners has discarded.
References
- Multistep Synthesis of Paracetamol in Continuous Flow
- Convergent approach for direct cross-coupling enabled by flash irreversible generation of cationic and anionic species
- Completeness and Diversity in Depth-First Proof-Number Search with Applications to Retrosynthesis
- Diverse and feasible retrosynthesis using GFlowNets
- RETRO SYNFLOW: Discrete Flow-Matching for Accurate and Diverse Single-Step Retrosynthesis
- When Single Answer Is Not Enough: Rethinking Single-Step Retrosynthesis Benchmarks for LLMs
- TotSyn: A Total Synthesis Reaction Dataset for Machine Learning in Organic Chemistry
- DARWIN: A Framework for Target Specific Diversity Constrained Natural Product Like Molecule Generation
- Pushing Biomolecular Utility-Diversity Frontiers with Supergroup Relative Policy Optimization
- AlphaSAGE: Structure-Aware Alpha Mining via GFlowNets for Robust Exploration