Quantum ChemistryQuantum Chemistry

Instant Quantum Chemistry for Small Molecules: The 2026 Guide

The new generation of neural network potentials makes DFT-quality quantum chemistry calculations instant and free.

SciRouter Team
April 11, 2026
12 min read

Five years ago, if a medicinal chemist asked a computational colleague for the pKa of a new compound, the answer came back in hours or days. Today it comes back in under a second, from a neural network trained on millions of DFT calculations, through a single API call. This is the quietly revolutionary part of modern drug discovery tooling, and this post is a practical guide to what it means for you.

What instant quantum chemistry is

For decades, quantum chemistry meant one thing: density functional theory or more expensive wavefunction methods, running on clusters, taking hours to days per molecule. The accuracy was excellent. The speed was terrible.

Neural network potentials change this by learning from large collections of DFT calculations. Once trained, they evaluate a new molecule in milliseconds, producing energies and forces that are close to what the underlying DFT would give you. Because everything a chemist wants (pKa, logP, BDE, reactivity) derives from those energies, everything inherits the same speed.

The end result is that a single HTTP call can now return the same information that used to require a full quantum chemistry pipeline.

What you can ask for

A representative small-molecule quantum chemistry pipeline now exposes the following properties through one API:

Energies and forces

The direct outputs of the neural network. Usually you do not look at them directly, but they power everything downstream.

Optimized geometries

Relax a SMILES-generated structure to a local energy minimum in under a second. The result is a 3D geometry of the same quality as DFT, ready to feed into docking, MD, or visualization.

pKa and ionization state

pKa for every ionizable site, plus the dominant microstate at a given pH. See our pKa predictor comparison for how this compares to empirical tools.

logP and TPSA

logP from thermodynamic cycles and TPSA from the relaxed 3D structure. Both are fundamental drug-like property descriptors.

Bond dissociation energies

BDE for each bond, ranked from weakest to strongest. The weakest C-H bonds are often the first targets for oxidative metabolism.

Fukui reactivity indices

Per-atom scores for nucleophilic and electrophilic attack, useful for reasoning about selectivity in reactions and metabolism.

Atomic charges

Partial charges from the neural network, consistent with the energy. Useful for electrostatic surface analysis and as input to downstream force field fits.

How it compares to DFT

For the property categories above, on drug-like molecules, accuracy typically falls within the spread you would see between different DFT functionals. That means:

  • pKa agreement of 0.4 to 0.6 pKa units.
  • logP agreement of 0.4 log units.
  • BDE agreement of a few kilocalories per mole.
  • Relaxed geometry RMSD of fractions of an Angstrom.

Experimental values for these same quantities typically scatter by a similar amount when measured by different techniques. Neural network potentials are therefore often within experimental noise, which is the best you can ask for.

Note
The big speed gap between neural network potentials and DFT means you can now compute quantum-grade properties on every molecule in a library instead of only the handful you could afford before. This is a quantitative change that becomes a qualitative one at scale.

When to still use DFT

Neural network potentials are powerful but not universal. Use traditional DFT when:

  • You need reaction pathways and transition states for chemistries that are not well represented in neural network training data.
  • You are working with transition metal complexes or multiconfigurational electronic structure.
  • Your molecule is significantly larger than the typical drug-like size range (more than about 100 heavy atoms).
  • You are asking a question that depends on very small energy differences, for example tight thermodynamic integration.

For everything else, neural network potentials are the practical default in 2026.

How to integrate into your workflow

A few patterns that work well:

  • Library triage. Compute logP, TPSA, and pKa for every compound in your virtual library and filter by drug-likeness before any expensive downstream work.
  • Protonation state preparation. Determine the dominant microspecies at physiological pH and feed that state to docking, free energy calculations, or structure prediction rather than the as-drawn form.
  • Metabolic stability triage. Rank BDE across the molecule and flag compounds with unusually weak C-H bonds as metabolic liabilities.
  • Agent tool calls. Expose the API as a tool for LLM agents so they can answer quantitative chemistry questions with real numbers.

For code-level examples, see our developer guide which walks through curl and Python calls with full examples.

A brief taxonomy of the models

Two names dominate small-molecule neural network quantum chemistry in 2026:

  • Egret-1 — DFT-quality energy model specialized for organic chemistry. Strong on pKa, logP, and BDE. Covered in our Egret-1 overview.
  • AIMNet2 — neural network potential for organic chemistry with explicit charge support. Strong on conformer optimization and reactions. Covered in our AIMNet2 guide.

Both are available through the SciRouter quantum chemistry properties tool. For most users the choice is transparent because the endpoint picks whichever model is most accurate for the requested property.

What this unlocks

Instant quantum chemistry is one of those capabilities that sounds incremental until you try it. The first time you put 10,000 molecules through a pKa calculation in 20 minutes, you realize the whole shape of what is possible has changed. Virtual screens that used to be filtered by crude fingerprint descriptors can now be filtered by quantum-grade properties. LLM agents that used to guess logP can now compute it. Medicinal chemists can now ask “what is the dominant tautomer” and get a real answer in the time it takes to sip coffee.

Bottom line

Neural network potentials have turned quantum chemistry from an expensive, slow, specialist tool into a routine screening step. If you are doing any computational work with small molecules in 2026, this is one of the most immediately useful capabilities you can add to your stack.

Open the QM properties tool →

Frequently Asked Questions

What does 'instant quantum chemistry' actually mean?

It means that the properties you used to get from a DFT calculation, which takes minutes to hours, are now available in under a second from a neural network potential trained on DFT data. The accuracy is close to the underlying DFT, but the speed is a million times faster.

Which properties can I get instantly?

Energies, forces, relaxed geometries, pKa, logP, TPSA, bond dissociation energies, Fukui reactivity indices, and atomic charges. Anything that reduces to energies and forces on a small molecule is fair game.

Is this accurate enough to replace DFT?

For the properties above, on drug-like molecules, yes, within the noise of DFT itself. For reactive chemistry, exotic electronic structure, or transition metal complexes, DFT is still the right tool. For the 90 percent of medicinal chemistry workflows that do not need those things, neural network potentials are the new default.

Do I need to know anything about quantum chemistry to use these tools?

Not really. The APIs take a SMILES string and return a JSON object with property values. You do not need to understand basis sets, exchange-correlation functionals, or integration grids. That said, knowing roughly what each property means helps you interpret the results sensibly.

What is the catch?

The main catch is that neural network potentials can fail silently on molecules that are far from their training distribution. Unusual functional groups, rare tautomers, or exotic scaffolds may give answers that look reasonable but are quantitatively wrong. Spot-check with a second method on novel chemistry.

Where should I start?

Open the QM properties tool on SciRouter, paste in a SMILES string, and look at the JSON response. Once you see what comes back, you can wire it into whatever downstream workflow you care about.

Try this yourself

500 free credits. No credit card required.