A System-Independent Metadynamics Strategy for Reactive Training Data: Application to Gas-Phase Organic Reactions
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.
Kai: We're looking at the paper: "A System-Independent Metadynamics Strategy for Reactive Training Data".
Mira: Here's the gist of the paper. General-purpose machine-learning interatomic potentials (MLIPs) for organic reactions need to be accurate on both the minimum energy path (MEP) for static evaluation of basic properties and the broader configurational space for simulating reaction dynamics.
Kai: The core methodological contribution is a system-independent collective variable built from Cartesian RMSD against a dynamically expanding library of time-averaged references, with local-domain partitioning for tunable spatial resolution.
Mira: Applied within a DPGEN2 concurrent-learning workflow, this produces OpenRxn26, a dataset of 1.8 M DFT-labeled configurations covering reactions from neutral singlet unimolecular reactants in the H/C/N/O chemical space with Nheavy ≤ 30.
Kai: A DPA3 model trained on OpenRxn26 (denoted DPA3 rxn) serves as the production-grade demonstration to evaluate the data-model pipeline The Sampling Strategy The core of the methodology is a spatiotemporally resolved, system-independent collective variable: Cartesian RMSD within randomly partitioned local domains against an expanding list of time-averaged reference geometries.
Title and authors: Kai: So we're talking about this paper today, "A System-Independent Metadynamics Strategy for Reactive Training Data: Application to Gas-Phase Organic Reactions." It seems like they're tackling a fundamental problem in machine learning potentials.
Mira: Right, it addresses the need for accurate models that work both for static property checks along the minimum energy path and also for simulating how reactions actually happen in time, which is a big gap right now.
Lev: I'm wondering what they mean by system-independent when they talk about this approach. Does that mean it works well regardless of the specific chemical reaction we are looking at?
Kai: Exactly, Lev. The core idea is using a collective variable built from Cartesian root mean square deviation against an expanding library of time averaged reference geometries, which is designed to be system-independent >
Mira: That sounds like they're creating a way to explore the entire chemical space without being locked into one specific reaction pathway from the start, which seems smart for general training data >
Lev: If you are trying to run this on real hardware, how does that collective variable strategy translate into actual computational overhead when you are sampling these complex geometries?
Kai: The paper says they use metadynamics as the main engine for exploration, but they supplement that with structural relaxation towards a transition state to get better coverage around those important points >
Mira: And then they build the bias potential by summing up Gaussian hills where each hill is centered on a collective variable value that was visited earlier in the trajectory, which is how you accelerate finding rare events >
Lev: So it’s using past exploration paths to guide future searches, trying to sample both chemical and conformational spaces at once, but does that mean the convergence speed stays good when the reaction involves a big change in geometry?
Kai: It's designed for that reason, Lev. The way they partition the local domains allows them to tune the spatial resolution, so you can look at things on a very small scale or cover a wider region >
Mira: They actually built this entire pipeline using a concurrent learning workflow called DPGEN2, which cycles through model training and exploration and first-principles labeling all together >
Lev: Concurrent learning sounds like a lot of juggling for the AI system. How do they decide which configurations are good enough to get labeled by DFT in the first place?
Kai: They select candidates for labeling based on two trust levels, sigma low and sigma high, which are determined by how much deviation there is from the model's prediction >
Mira: That uncertainty measure is quantified by an epsilon value, which helps filter out the configurations that aren't representative of what the current model understands well enough >
Lev: So it’s a feedback loop where the AI tries to learn, explores with metadynamics, and then uses its own confidence to ask for high-level verification from first principles calculations >
Kai: And this process generates OpenRxn26, which is a dataset of one point eight million DFT labeled configurations covering neutral singlet unimolecular reactions in the H/C/N/O chemical space with N heavy atoms less than or equal to thirty >
Mira: That is a huge dataset for organic chemistry because it covers environments that you just don't find in the existing community datasets >
Lev: I see. But if we look at what this means for running on real hardware, the paper mentions training a DPA3 model on this OpenRxn26 data, which then becomes their production-grade demonstration >
Kai: The DPA3 model is the final product they test, and they found that when you train it on OpenRxn26, it shows over an order of magnitude force mean absolute error inflation in regions that are distinguished by this new dataset >
Mira: That's significant because if a model trained on something else, like OMol25, shows that much error in these specific reactive areas of OpenRxn26, it tells us the existing datasets are missing crucial information about how things change during a reaction >
Lev: So the gain here is that this new data allows for training models that are much more accurate on those specific reactive environments, which is what I need for real hardware where you can't afford huge errors >
Kai: And they also claim that the DPA3 rxn model is the only one in their benchmark suite to reach a one point zero kcal per mole energy accuracy on off-minimum energy path reactive trajectories when compared to the labeling method >
Mira: That’s important because it shows that relying only on static analysis or just anchoring everything to the minimum energy path isn't enough for guaranteeing reliability when you are simulating dynamics >
Lev: So, they found that quasi-static sampling and MEP anchored benchmarks aren't sufficient for reliable dynamics, which is a tough point to make when you want to build something robust >
Kai: And they show that DPA3 rxn achieves transferable accuracy on barrier heights and reaction energies across a representative set of test reactions, which is what we need for practical application >
Mira: They also noted that DPA3 rxn has the smallest trajectory-to-trajectory variation across diverse reactions compared to the labeling methods they tested, which suggests it’s very consistent >
Lev: So, if we look at what this means for simulation, OpenRxn26 is positioned as MD ready reactive training data for gas phase neutral singlet organic reactions >
Kai: It’s a lot of data that covers atomic environments that general purpose datasets like OMol25 simply do not cover, which really expands the scope of what we can simulate >
Mira: The main thing they prove is that this spatiotemporally resolved collective variable strategy is efficient and general-purpose enough to generate high quality training data for MLIPs >
Lev: I just want to reiterate the limitation they mentioned: since it’s driven by sampling against an expanding list of references, there might be some uncertainty about how well it captures every possible reactive event, even though they mitigate that with the concurrent learning approach >
Kai: That uncertainty is managed by letting the exploration engine handle the discovery while the training signal builds on that discovered space >
Mira: So overall, OpenRxn26 provides a very robust source of data for simulating these types of reactions dynamically, which is what we need to move beyond just static evaluations >
Lev: It’s a solid piece of work because it shows how to generate high quality reactive data efficiently, and that the DPA3 model is actually performing well on the dynamics side when trained this way >
Kai: Yeah, so OpenRxn26 is definitely something to look at if you're working on building models for complex organic reactions in a dynamic setting >
The paper's summary: Kai: So, this paper is about building a way to generate training data for machine learning potentials that works across different types of reactions, not just one specific reaction path.
Mira: Right, they've put together a whole system-independent method using something called metadynamics to explore the reaction space.
Lev: I mean it uses this collective variable that is built from the distance between atoms and a set of reference geometries, which is supposed to work whether you're looking at a small bond change or a big conformational shift.
Kai: Exactly, it’s spatiotemporally resolved, meaning you can track what happens in space and time simultaneously, and they use this collective variable to drive the metadynamics sampling.
Mira: They’re using that sampling to build this dataset called OpenRxn26 which has eighteen million configurations labeled by density functional theory calculations.
Lev: That's a lot of DFT work, but the point is they are covering chemical spaces—the H, C, N, and O elements—and they kept the heavy atoms under thirty.
Kai: And then they train this DPA3 model on that OpenRxn26 data to make it ready for real use in simulating dynamics.
Mira: The big result they show is that this DPA3 model is remarkably accurate, especially when it comes to barrier heights and reaction energies across different test reactions.
Lev: They say this is important because DPA3 rxn is the only model in their set that hits one point zero kcal per mole energy accuracy on trajectories that aren't just along the minimum energy path.
Kai: That means you can use it for actual molecular dynamics simulations, not just looking at static pictures of a reaction.
Mira: And they also showed that this data is robust; even when you test the model against geometries from other datasets, like OMol25, the error inflation in those reactive areas is massive.
Lev: That tells us that whatever information these existing datasets are missing about how things change during a reaction is really significant.
Kai: It proves this new sampling strategy is general-purpose; it isn't just tailored to one reaction type.
Mira: So, the whole point here is that OpenRxn26 becomes this MD-ready training data for gas phase organic reactions, covering environments those other datasets just don't touch.
Lev: It’s a solid foundation for building models that can handle the actual movement of atoms during a chemical transformation.
Kai: And we’re going to look at how they actually built this OpenRxn26 dataset in their next segment, starting with the concurrent learning workflow they used.
The paper's improvements: Mira: So, moving past how they built the data, Kai and I want to talk about what this paper suggests as improvements to this entire reactive training data pipeline.
Kai: Right, they’re looking at how we can make this sampling strategy even better for future work.
Lev: What are they suggesting about the core collective variable approach? Is it too limited in its spatial resolution?
Kai: They suggest that the system-independent way they built that collective variable, using Cartesian root mean square deviation against those time averaged reference geometries, can be used to explore a much wider range of scales.
Mira: They want to get from just exploring local bond changes to covering really diffuse motion events too.
Lev: So they’re trying to make sure the method can handle everything from tiny vibrations up to big conformational shifts in the molecule.
Kai: Exactly, and they frame it as a way to tune the spatial resolution on the fly, so you can cover smaller or larger regions depending on what you need.
Mira: They also focus on how they integrate that concurrent learning workflow more deeply into the exploration process.
Lev: Is there a specific tweak to how the model training and first principles labeling cycle together?
Kai: They emphasize making that cycle tighter, where the AI system automatically selects configurations for DFT labeling based on its own internal uncertainty.
Mira: So instead of just running one thing after another, they want a more integrated feedback loop where the model learns, explores with metadynamics, and then asks for verification from first principles all at once.
Lev: That sounds like it could reduce the overhead because you're not doing separate steps for each phase.
Kai: It should help speed up the entire process of getting high-quality labeled data without needing huge amounts of extra computation just to verify points.
Mira: The implication is that future models trained this way will be more robust because they’ll have been vetted by the AI itself before we even get to the expensive DFT step.
Lev: It means less wasted time on labeling configurations that are obviously bad guesses from the model's perspective, which lowers the overall cost of building a reliable potential.
Kai: So, they're pushing for a smarter way for the AI to guide its own exploration and verification process.
Mira: That’s what it is; they want to move towards systems that are self-correcting in their data curation pipeline rather than just relying on external human guidance at every step.
Lev: It makes sense because running these massive sampling simulations takes so much time, and any efficiency gain in the labeling phase really helps when you're building something for real hardware.
Kai: So the next thing we look at is how this data actually performs when we try to run models trained on it in practice.
Conclusion: Kai: So we’ve covered how they built this method and what the results show regarding accuracy on those off minimum energy path trajectories for the paper "A System-Independent Metadynamics Strategy for Reactive Training Data: Application to Gas-Phase Organic Reactions."
Mira: Yeah, essentially, they've given us a really solid source of training data, OpenRxn26.
Lev: I mean that this dataset is MD ready because DPA3 rxn achieved one point zero kcal per mole energy accuracy on dynamics derived configurations.
Kai: That means we can actually use this for simulations, not just static property checks.
Mira: It shows that the way they sampled the space works generally across different chemical systems, which is what makes it useful for broad applications.
Lev: The caveat is that they acknowledge this method relies on a spatiotemporal collective variable built from RMSD against an expanding reference library, so we still have to trust how well it captures every single reactive event in the whole space.
Kai: It’s a trade-off, right? A lot of coverage versus absolute certainty on every tiny detail.
Mira: Exactly, and that uncertainty is managed by their concurrent learning workflow which filters configurations based on model deviation.
Lev: From an engineering standpoint, it means we can use this data to train models that are robust even when the training set isn't perfectly aligned with the exact dynamics we want to simulate.
Kai: It’s a big step forward for moving MLIPs out of just static analysis and into real dynamic simulations.
Mira: And they’ve set a high bar, showing that this approach is efficient enough to generate eighteen million labeled configurations in the H, C, N, and O chemical space.
Lev: It’s a lot of DFT labeling work, but the payoff is a data source that seems reliable across diverse reactions.
Kai: So for anyone working on building models for complex organic reactions in a dynamic setting, OpenRxn26 provides this very robust training foundation.
Mira: That's the core implication: we have a better way to generate high-quality reactive training data without being locked into just one reaction mechanism.
Lev: It’s certainly something to keep an eye on as we try to build the next generation of quantum error correction schemes that need reliable simulation benchmarks.
Kai: Alright, that wraps up our look at this paper, "A System-Independent Metadynamics Strategy for Reactive Training Data: Application to Gas-Phase Organic Reactions."
Mira: Thanks for tuning in. We’ll be back soon with another paper from arXiv.
AI for Science Institute of Beijing 100080, P. R. China · School of Artificial Intelligence and Data Science, State Key Laboratory of Precision and Intelligent Chemistry, University of Science and Technology of China, Hefei 230026, P. R. China · Suzhou Institute for Advanced Research, University of Science and Technology of China, Suzhou 215123, P. R. China · Suzhou Big Data & AI Research and Engineering Center, Suzhou 215123, P. R. China · State Key Laboratory of Coordination Chemistry, Key Laboratory of Mesoscopic Chemistry of Ministry of Education, Nanjing University, Nanjing 210093, P. R. China · The Institute of Green Chemistry and Engineering, Nanjing University, Suzhou 215163, P. R. China · Shanghai Engineering Research Center of Molecular Therapeutics and New Drug Development, School of Chemistry and Molecular Engineering, East China Normal University, Shanghai 200062, P. R. China · Shanghai Innovation Institute of Shanghai 200003, P. R. China · National Key Laboratory of Computational Physics, Institute of Applied Physics and Computational Mathematics, Fenghao East Road 2, Beijing 10094, P. R. China · HEDPS CAPT College of Engineering, Peking University, Beijing 100871, P. R. China
physics.chem-ph
Submitted: 2026-09-24
Updated: 2026-11-12
Code: https://github.com/Vibsteamer/OpenRxn26_benchmark-ref_sampling-script
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 100/100
The gist: General-purpose machine-learning interatomic potentials (MLIPs) for organic reactions need to be accurate on both the minimum energy path (MEP) for static evaluation of basic properties and the…
Key concepts
- System-Independent Collective Variable (CV)
- This CV measures the distance between a configuration and a growing library of reference geometries, calculated locally within specific parts of the molecule. This allows the simulation to explore different spatial scales simultaneously, from small local changes to large structural movements, making it robust for general organic reactions.
- Concurrent Learning Workflow (DPGEN2)
- This is an iterative process where a machine learning model is trained on existing and newly collected data. It cycles through training the model, exploring the potential energy surface using metadynamics, and then using the model's uncertainty to select configurations for high-accuracy first-principles labeling.
- Transferable Accuracy
- This refers to a model's ability to accurately predict reaction energies and barrier heights across different types of reactions. DPA3 rxn proved it could maintain high accuracy (1.0 kcal/mol MAE) even when tested on dynamic trajectories, proving it is reliable for real molecular dynamics simulations rather than just static snapshots.
Terminology
Summary
General-purpose machine-learning interatomic potentials (MLIPs) for organic reactions need to be accurate on both the minimum energy path (MEP) for static evaluation of basic properties and the broader configurational space for simulating reaction dynamics. p. 1 Abstract
General-purpose machine-learning interatomic potentials (MLIPs) for organic reactions need to be accurate on both the minimum energy path (MEP) for static evaluation
of basic properties and the broader configurational space for simulating reaction dynamics. Existing general datasets for gas-phase organic reactions rely on quasi-static
relaxation that confines configurations to the MEP vicinity, so models trained on
them could fail on direct molecular-dynamics trajectories; the gap is methodological,
not a question of dataset size. This paper presents a sampling methodology that produces reactive training data with
sufficient off-MEP coverage to make general MLIPs MD-ready for gas-phase organic reactions. The core is a spatiotemporally resolved, system-independent CV: Cartesian RMSD within
randomly partitioned local domains (spatial resolution) against an expanding list of timeaveraged reference geometries (temporal noise filtering). The
CV drives metadynamics as the main exploration engine, supplemented by structural
relaxation towards transition state (TS) to augment the coverage around TS. Within a
concurrent-learning workflow, this produces OpenRxn26, a dataset of 1. The
CV drives metadynamics as the main exploration engine, supplemented by structural
relaxation towards transition state (TS) to augment the coverage around TS. Within a
concurrent-learning workflow, this produces OpenRxn26, a dataset of 1. The outcome
is OpenRxn26, a dataset of 1. 8 M DFT-labeled configurations covering reactions from
neutral singlet unimolecular reactants in the H/C/N/O chemical space with Nheavy ≤ 30,
containing reactive atomic environments underrepresented in community datasets. The outcome
is OpenRxn26, a dataset of 1. 8 M DFT-labeled configurations covering reactions from
neutral singlet unimolecular reactants in the H/C/N/O chemical space with Nheavy ≤ 30,
containing reactive atomic environments underrepresented in community datasets.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
The Sampling Strategy
The core of the methodology is a spatiotemporally resolved, system-independent collective variable: Cartesian RMSD within randomly partitioned local domains against an expanding list of time-averaged reference geometries. This CV drives metadynamics as the main exploration engine, supplemented by structural relaxation towards transition state (TS) to augment the coverage around TS. p. 1 We introduce a spatiotemporally resolved, systemindependent collective variable (CV): Cartesian RMSD within randomly partitioned
local domains against an expanding list of time-averaged reference geometries. The
CV drives metadynamics as the main exploration engine, supplemented by structural
relaxation towards transition state (TS) to augment the coverage around TS. We introduce a spatiotemporally resolved, systemindependent collective variable (CV): Cartesian RMSD within randomly partitioned
local domains against an expanding list of time-averaged reference geometries. The
CV drives metadynamics as the main exploration engine, supplemented by structural
relaxation towards transition state (TS) to augment the coverage around TS. 1 Sampling strategy
In metadynamics,50,51 MD trajectories propagate on a PES augmented by an accumulating
sum of Gaussian bias kernels (hills); each hill is centered at a collective variable (CV)
value visited earlier in the trajectory, so the bias progressively raises the currently occupied
free-energy basin until the system escapes over a barrier in CV space. This mechanism
not only accelerates rare reactive events but also samples chemical and conformational
spaces simultaneously, making metadynamics well-suited as the engine of a general-purpose
reactive-data curation workflow. 1 Sampling strategy
In metadynamics,50,51 MD trajectories propagate on a PES augmented by an accumulating
sum of Gaussian bias kernels (hills); each hill is centered at a collective variable (CV)
value visited earlier in the trajectory, so the bias progressively raises the currently occupied
free-energy basin until the system escapes over a barrier in CV space. This mechanism
not only accelerates rare reactive events but also samples chemical and conformational
spaces simultaneously, making metadynamics well-suited as the engine of a general-purpose
reactive-data curation workflow. 1 Sampling strategy
In metadynamics,50,51 MD trajectories propagate on a PES augmented by an accumulating
sum of Gaussian bias kernels (hills); each hill is centered at a collective variable (CV)
value visited earlier in the trajectory, so the bias progressively raises the currently occupied
free-energy basin until the system escapes over a barrier in CV space. This mechanism
not only accelerates rare reactive events but also samples chemical and conformational
spaces simultaneously, making metadynamics well-suited as the engine of a general-purpose
reactive-data curation workflow.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
The Training Signal is Built
The training signal is built through a concurrent learning workflow called DPGEN2. This iterative scheme cycles through three steps: model training, exploration, and first-principles labeling. p. 8 c, DPGEN2 concurrent
learning. Each iteration cycles through training (an ensemble of MLIPs is fine-tuned on
the current dataset), exploration (sampling strategy in a, runs on PES of one MLIP in the
ensemble; sampled candidates are classified by ensemble deviation ϵ into accurate / candidate
/ failed), and labeling (selected candidates labeled by DFT and added to the training set). c, DPGEN2 concurrent
learning. Each iteration cycles through training (an ensemble of MLIPs is fine-tuned on
the current dataset), exploration (sampling strategy in a, runs on PES of one MLIP in the
ensemble; sampled candidates are classified by ensemble deviation ϵ into accurate / candidate
/ failed), and labeling (selected candidates labeled by DFT and added to the training set). Scanning the partition radius captures
reactions over a range of spatial scales, from local bond-changing to diffuse-motion-gated
events. Selected snapshots from reactive windows seed TS optimizations (dimer method);
metadynamics seeds and sparse configurations along TS-relaxation trajectories are passed to
the uncertainty evaluation in exploration step of DPGEN2. c, DPGEN2 concurrent
learning. Each iteration cycles through training (an ensemble of MLIPs is fine-tuned on
the current dataset), exploration (sampling strategy in a, runs on PES of one MLIP in the
ensemble; sampled candidates are classified by ensemble deviation ϵ into accurate / candidate
/ failed), and labeling (selected candidates labeled by DFT and added to the training set).Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the Gains Come From
The gains come from the ability of OpenRxn26 to cover atomic environments that general-purpose datasets such as OMol25 do not, with a model trained on OMol25 showing more than an order of magnitude force-MAE inflation on the OpenRxn26-distinguished regions. Furthermore, DPA3 rxn is the only model in the benchmark suite to reach 1. p. 1 Trained on OpenRxn26, a DPA3 model (denoted DPA3 rxn)
achieves transferable accuracy on barrier heights and reaction energies. On off-MEP
reactive trajectories, DPA3 rxn is the only model in the benchmark suite to reach 1. 0
kcal/mol energy accuracy compared with the labeling method, where the domain MLIP
leading on static benchmarks degrades several-fold (e. MACE OMol25), showing the
insufficiency of quasi-static sampling and MEP-anchored benchmarks for guaranteeing
dynamics reliability of MLIPs.Where the paper says it
Where the paper says it
Transferable Accuracy and Reliability
DPA3 rxn achieves transferable accuracy on barrier heights and reaction energies across representative test reactions. In the benchmark suite, DPA3 rxn is the only model to deliver 1. p. 1 Trained on OpenRxn26, a DPA3 model (denoted DPA3 rxn)
achieves transferable accuracy on barrier heights and reaction energies. On off-MEP
reactive trajectories, DPA3 rxn is the only model in the benchmark suite to reach 1. Trained on OpenRxn26, a DPA3 model (denoted DPA3 rxn)
achieves transferable accuracy on barrier heights and reaction energies. On off-MEP
reactive trajectories, DPA3 rxn is the only model in the benchmark suite to reach 1. 8 kcal/mol lower
than the unmodified reaction, the trajectory traverses the TS region 50% more slowly because
long-range C60 interactions induce complex motions that hinder reactant alignment. 25
Addressing these challenges requires PES descriptions that are accurate and efficient,
covering not only the MEP but also the configurational space sampled by reaction dynamics. MACE OMol25), showing the
insufficiency of quasi-static sampling and MEP-anchored benchmarks for guaranteeing
dynamics reliability of MLIPs. OpenRxn26 thus provides MD-ready reactive training data for gas-phase neutral singlet organic reactions, verifying the generality and
efficiency of the sampling strategy. MACE OMol25), showing the
insufficiency of quasi-static sampling and MEP-anchored benchmarks for guaranteeing
dynamics reliability of MLIPs. OpenRxn26 thus provides MD-ready reactive training data for gas-phase neutral singlet organic reactions, verifying the generality and
efficiency of the sampling strategy. Section 2 describes the methodology,
including the sampling strategy, concurrent-learning workflow, CV formulation, and training
protocol of DPA3 models. Section 3 presents results benchmarking the atomic-environment
coverage of OpenRxn26, the transferable accuracy of DPA3 rxn on basic reaction properties,
and DPA3 rxn’s reliability on reactive dynamics trajectories.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Conclusion
OpenRxn26 thus provides MD-ready reactive training data for gas-phase neutral singlet organic reactions, verifying the generality and efficiency of the sampling strategy. OpenRxn26 covers atomic environments that general-purpose datasets such as OMol25 do not, with a model trained on OMol25 showing more than an order of magnitude force-MAE inflation on the OpenRxn26-distinguished regions. p. 1 8 M DFT-labeled
configurations covering neutral singlet unimolecular reactions in the H/C/N/O chemical
space (Nheavy ≤ 30), containing reactive atomic environments underrepresented in
community datasets. Trained on OpenRxn26, a DPA3 model (denoted DPA3 rxn)
achieves transferable accuracy on barrier heights and reaction energies. Trained on OpenRxn26, a DPA3 model (denoted DPA3 rxn)
achieves transferable accuracy on barrier heights and reaction energies. On off-MEP
reactive trajectories, DPA3 rxn is the only model in the benchmark suite to reach 1. 0 kcal/mol overall energy accuracy with respect to the labeling method of training data,
outperforming other tested models. OpenRxn26 thus provides MD-ready reactive training
data, establishing the metadynamics-based sampling strategy as an efficient data solution for
building general-purpose MLIPs on gas-phase organic reactions. 0 kcal/mol overall energy accuracy with respect to the labeling method of training data,
outperforming other tested models. OpenRxn26 thus provides MD-ready reactive training
data, establishing the metadynamics-based sampling strategy as an efficient data solution for
building general-purpose MLIPs on gas-phase organic reactions. MACE OMol25), showing the
insufficiency of quasi-static sampling and MEP-anchored benchmarks for guaranteeing
dynamics reliability of MLIPs. OpenRxn26 thus provides MD-ready reactive training data for gas-phase neutral singlet organic reactions, verifying the generality and
efficiency of the sampling strategy.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 1 ---
A System-Independent Metadynamics Strategy
for Reactive Training Data: Application to
Gas-Phase Organic Reactions
Pruning redundant reasoning tokens
General-purpose machine-learning interatomic potentials (MLIPs) for organic reactions need to be accurate on both the minimum energy path (MEP) for static evaluation of basic properties and the broader configurational space for simulating reaction dynamics. The core methodological contribution is a spatiotemporally resolved, system-independent collective variable built from Cartesian RMSD against an expanding list of time-averaged reference geometries. p. 1 Existing general datasets for gas-phase organic reactions rely on quasi-static
relaxation that confines configurations to the MEP vicinity, so models trained on
them could fail on direct molecular-dynamics trajectories; the gap is methodological,
not a question of dataset size. We introduce a spatiotemporally resolved, systemindependent collective variable (CV): Cartesian RMSD within randomly partitioned
local domains against an expanding list of time-averaged reference geometries.Where the paper says it
The core of the methodology is a spatiotemporally resolved, system-independent collective variable: Cartesian RMSD within randomly partitioned local domains against an expanding list of time-averaged reference geometries. This CV drives metadynamics as the main exploration engine, supplemented by structural relaxation towards transition state (TS) to augment the coverage around TS. The bias potential is constructed as a sum of Gaussian hills where each hill is centered at a collective variable (CV) value visited earlier in the trajectory.
DPA3 rxn achieves transferable accuracy on barrier heights and reaction energies across representative test reactions. In the benchmark suite, DPA3 rxn is the only model to deliver 1.0 kcal/mol overall MAE on dynamics-derived configurations and the smallest trajectory-to-trajectory variation across diverse reactions with respect to labeling methods of training data. This establishes OpenRxn26 as MD-ready reactive training data for gas-phase neutral singlet organic reactions.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Spatiotemporally resolved collective variable CV: The system can use a CV built from
Cartesian RMSD within randomly partitioned local domains against an expanding list of time-averaged reference geometries
to drive metadynamics, allowing for exploration across a range of spatial scales fromlocal bond-changing to diffuse-motion-gated events.
-
Concurrent learning workflow: The AI system can integrate a
concurrent-learning workflow
where it cycles throughmodel training, exploration, and first-principles labeling,
automatically selecting configurations for DFT labeling based on the model's uncertainty (quantified by ensemble disagreement epsilon). -
Model generalization: The trained DPA3 model can achieve
transferable accuracy on barrier heights and reaction energies
even onoff-MEP reactive trajectories,
making it suitable for simulating direct molecular dynamics trajectories rather than just static evaluations.
Abstract
General-purpose machine-learning interatomic potentials (MLIPs) for organic reactions need to accurately describe both minimum-energy paths (MEPs) for evaluating reaction energetics and the broader configurational space for simulating reaction dynamics. Existing general datasets for gas-phase reactions largely rely on quasi-static optimization near MEPs, leaving off-MEP regions under-covered and model reliability in reactive molecular dynamics weakly supported; the gap is methodological, not a question of dataset size. We introduce a spatiotemporally resolved, system-independent collective variable (CV): Cartesian RMSD within randomly partitioned local domains against an expanding list of time-averaged reference geometries. The CV drives metadynamics as the exploration engine, supplemented by transition state (TS) optimization to augment coverage around TSs. Within a concurrent-learning workflow, this produces OpenRxn26, a dataset of 1.8 M DFT-labeled configurations covering neutral singlet unimolecular reactions in the H/C/N/O chemical space (N heavy at most 30), containing reactive atomic environments underrepresented in community datasets. Trained on OpenRxn26, a DPA4 model (DPA4_rxn26) is the only model in the benchmark suite to achieve transferable sub-kcal/mol energy accuracy on barrier heights, reaction energies and off-MEP reactive trajectories, relative to the training DFT reference, demonstrating the efficiency and generality of the sampling strategy. Domain MLIPs performing well on static benchmarks degrade several-fold on off-MEP reactive trajectories (e.g. MACE-OMol25), showing the insufficiency of MEP-anchored benchmarks for guaranteeing dynamics reliability. The sampling strategy enables efficient development of reliable general-purpose reactive MLIPs, with OpenRxn26 as a demonstration of MD-ready training data for gas-phase neutral singlet organic reactions.
Sources
- The Open Molecules 2025 (OMol25) Dataset, Evaluations, and Models
- Cross Learning between Electronic Structure Theories for Unifying Molecular, Surface, and Inorganic Crystal Foundation Force Fields
Related papers
- A Machine Learning Model for the Chemistry of a Solvated Electron
- Compression of virtual spaces in transcorrelated methods via singular value decomposition: application to the G2 set
- Charge Transfer with a Spin. II: A Framework for Diabatization which Localizes Charge and Spin
- Diabatic Seam Space Sampling for Hydrogen Tunneling Systems with Nuclear-Electronic Orbital Theory
- Augmented Roothaan-Hall Hessian Applied to Spin-Restricted Open-Shell Density-Functional Theory
- The Same and Not the Same: A Machine-Checked Genealogy of Photochemical Theory