A quantum chemistry dataset containing ground-state and conical-intersection structures of 260k molecules

This paper introduces a comprehensive quantum chemistry dataset comprising ground-state and conical-intersection structures for 260,000 small molecules, calculated at the OM2/MRCI level, to facilitate the integration of photochemistry with machine learning for studying excited-state reaction processes.

The Gist

Imagine the world of molecules as a vast, hilly landscape. When a molecule absorbs light (like sunlight), it doesn't just sit still; it jumps up a hill into an "excited state." Usually, it wants to slide back down to its comfortable, resting spot (the ground state).

However, sometimes the landscape has a very special, tricky spot called a Conical Intersection (CI). Think of a CI as a magical funnel or a crossroads where two different hills merge into a single point. If a molecule rolls into this funnel, it can instantly switch tracks, changing its behavior completely. This is how things like photosynthesis work, how our eyes see light, or how some molecules protect themselves from getting damaged by the sun.

For a long time, scientists have been trying to map these funnels, but they've only been able to draw a few maps for specific, small towns. They couldn't build a global atlas because calculating these funnels is incredibly hard and slow.

What this paper does:
The researchers have built a massive digital atlas containing 260,000 different molecular "towns." For every single one, they mapped out:

1. The comfortable resting spot (the ground state).
2. The magical funnel where the tracks cross (the conical intersection).

How they built it:
To make this atlas, they used a clever shortcut. Imagine trying to draw a map of the entire world. If you tried to measure every single tree and rock with a laser (which is what "high-level" science usually does), it would take forever. Instead, these scientists used a "quick sketch" method (called OM2/MRCI). It's like using a fast, reliable drone to take photos of the landscape. It's not perfect down to the millimeter, but it's accurate enough to see the shape of the hills and where the funnels are. This speed allowed them to process a quarter of a million molecules.

The "Quality Control" Check:
Before publishing the atlas, they had to clean it up, just like a librarian organizing books:

• The "Broken Map" Check: Sometimes, when they tried to find the funnel, the molecule would fall apart (like a Lego castle collapsing). These broken pieces were thrown out because they aren't useful funnels; they're just debris.
• The "Wrong Address" Check: Sometimes, the math got confused and found a spot that looked like a funnel but was actually lower than the ground level (which is physically impossible). These were also removed.
• The Result: After throwing out the broken or confusing maps, they were left with a clean, usable dataset of about 260,000 molecules.

What's inside the dataset?
The dataset is like a giant library of molecular blueprints. It includes:

• The Shapes: The exact 3D coordinates of the atoms for both the resting state and the funnel state.
• The Energy: How much energy it takes to get to these spots.
• The Variety: The molecules are diverse. Some are simple chains, some are rings (like bicycle wheels), and some are complex fused structures. They are made of Carbon, Nitrogen, Oxygen, and Fluorine.

Why is this useful?
The authors say this dataset is a training ground for Artificial Intelligence (AI).
Think of it this way: If you want to teach a robot to recognize a funnel in a landscape, you can't just show it one picture. You need to show it millions of examples. This dataset provides those millions of examples. Now, AI can learn the patterns of where these funnels usually appear, helping scientists predict how new molecules might behave without having to do the slow, expensive calculations for every single one.

Important Note:
The authors are very clear: This is a qualitative tool. It's like a weather forecast that tells you "it might rain" or "it's sunny," which is great for planning a picnic or training a model. But if you need to build a skyscraper (a precise medical drug or a specific industrial chemical), you still need the "laser measurement" (high-level calculations) to get the exact details. This dataset is the map that guides you to the right neighborhood, not the blueprint for the house itself.

In short:
They built a massive, high-speed map of 260,000 molecular landscapes, highlighting the tricky "funnels" where chemical reactions happen. They cleaned the map, checked the details, and made it available so that AI can learn to predict these reactions faster than ever before.

Technical Summary

Technical Summary: The QCDGE-CI Dataset

Problem Statement
Conical intersections (CIs) are critical topological features on potential energy surfaces where electronic states of the same spin become degenerate, acting as "funnels" for nonadiabatic transitions in photoinduced reactions. Despite their fundamental importance in photophysics and photochemistry, comprehensive datasets containing CI structures and energies are scarce. Existing research is largely system-specific, relying on individual quantum chemistry calculations that do not scale well. While machine learning (ML) has advanced chemical research, its application to photochemistry remains limited by the lack of high-quality, large-scale quantum chemistry datasets specifically targeting CI properties. Most existing datasets focus on ground states or provide only preliminary excited-state information, hindering the development of scalable ML models for CI prediction.

Methodology
To address this data gap, the authors constructed the QCDGE-CI dataset, a quantum chemistry repository containing ground-state and S0-S1 conical-intersection structures for 260,541 small molecules. The dataset is derived from the QCDGE dataset but re-optimized to ensure consistency.

The construction workflow involved five primary steps:

1. Ground-State Geometry Optimization: Initial geometries from the QCDGE dataset were re-optimized at the semi-empirical OM2 level to ensure electronic structure consistency across the dataset.
2. Geometry Validation: A hybrid validation procedure was employed to ensure structural equivalence between the original and re-optimized geometries. This involved comparing InChI strings as a primary filter, followed by molecular graph comparisons for cases where InChI strings differed, to distinguish true structural changes from minor torsional variations.
3. Single-Point Calculations: After optimization, single-point energy calculations were performed at the OM2/MRCI(4,4) level. This method constructs a configuration-interaction expansion using three reference states (closed-shell, single HOMO-LUMO excitation, and double HOMO-LUMO excitation) to provide consistent energy estimates relative to the ground-state minimum.
4. Conical Intersection Optimization: S0-S1 CI structures were optimized using the Lagrange-Newton algorithm at the OM2/MRCI level. The ground-state geometries served as initial guesses without conformational sampling. To manage convergence issues inherent in large-scale optimizations, the maximum SCF iterations were set to 5,000.
5. Dataset Filtering and Quality Control: Several filtering steps were applied to the final dataset:
   • Fragmentation Check: Structures that dissociated into two or more disconnected components during optimization were removed, as they do not represent photochemical funnels.
   • Energy Ordering Check: Geometries where the CI energy was lower than the ground-state minimum (potentially indicating local minima convergence or connectivity changes) were excluded to avoid artifacts.

The OM2/MRCI method was selected for its balance between computational efficiency and reasonable accuracy in describing CIs, enabling the processing of over 260,000 molecules. The dataset focuses exclusively on S0-S1 CIs, which are computationally more manageable and relevant to many nonadiabatic processes.

Key Contributions and Results

• Dataset Scale and Scope: The final dataset comprises 260,541 molecules containing up to ten heavy atoms (C, N, O, F). It includes coordinates and energies for both ground-state minima and S0-S1 conical intersections.
• Chemical Diversity: Technical validation confirms the dataset's chemical diversity:
  • Elemental Composition: 14 distinct elemental compositions were identified, with molecules containing C, N, and O simultaneously constituting the largest category (~44.2%).
  • Compound Types: The dataset covers eight distinct compound types, including heterocycles (26.1%), fused heterocycles (25.1%), and heteroaromatics (14.4%).
  • Structural Features: Over 80% of molecules contain ring moieties. Principal moments of inertia (PMI) analysis reveals a wide range of molecular shapes, though CI geometries show a shift toward more three-dimensional architectures compared to ground states, likely due to bond breaking and torsion.
  • Functional Groups: 102 functional groups were detected, with C=C double bonds being the most abundant.
• Energy Distributions: The dataset exhibits a broad distribution of S1 vertical excitation energies (1–8 eV) and S0-S1 CI energies. Generally, CI energies are lower than the corresponding S1 vertical excitation energies at the ground-state minima.
• Data Availability: The dataset is provided in two formats: Final_property.hdf5 (binary file with geometries and energies) and final_all.csv (auxiliary file with identifiers and structural metadata). Python scripts for data extraction and technical validation are available via a GitHub repository.

Significance and Claims
The authors position the QCDGE-CI dataset as a resource to bridge the gap between photochemical insight and data-driven approaches. The primary significance lies in enabling large-scale pattern analysis and the training of machine learning models for predicting CI structures and energies.

The paper explicitly frames the dataset's utility with modesty:

• Qualitative to Semi-Quantitative: The dataset is intended for applications such as ML model training, preliminary screening, and large-scale analysis. It is not designed to replace high-level multi-reference calculations for individual systems.
• Methodological Limitations: The authors acknowledge the inherent approximations of the semi-empirical OM2/MRCI method and the challenges of CI optimization (e.g., convergence failures and fragmentation). However, they argue that despite these limitations, the dataset offers a practical foundation for exploring the vast chemical space of conical intersections, which was previously inaccessible due to computational costs.

By providing a standardized, large-scale resource, the authors aim to facilitate the development of AI tools that can accelerate the understanding of photochemical reaction mechanisms and aid in molecular design.