MaterialsMaterials Potentials

Machine Learning Potentials Are Replacing DFT in 2026

Why ML potentials like MACE and NequIP are replacing DFT for most materials science workflows. Speed, accuracy, and when you still need real DFT.

SciRouter Team
April 11, 2026
13 min read

For 30 years, density functional theory was the workhorse of computational materials science. You wanted a new cathode? DFT. You wanted a formation energy? DFT. You wanted to understand catalysis? DFT. The method had a Nobel Prize behind it and predictable accuracy, and most of the field's muscle memory was built around its quirks.

In 2026 that default is changing. Machine learning interatomic potentials (MLIPs) are fast enough, accurate enough, and universal enough that most production work is moving to them. DFT is still essential, but its role is narrowing. This post walks through what changed, where MLIPs now win, and where DFT is still the right tool.

The speed gap

A single-point DFT calculation on a 100-atom system with a plane-wave code and a modern functional takes minutes to hours on a CPU node. The same calculation with a modern MLIP like MACE-MP-0 takes a millisecond on the same hardware. Molecular dynamics with DFT forces is limited to thousands of steps. Molecular dynamics with MLIP forces can run billions.

This is not a small improvement. It is six orders of magnitude. It changes what is practical.

  • Before: Nanosecond ab-initio MD trajectories were major publications.
  • Now: Microsecond trajectories run overnight on a laptop.
  • Before: Screening 100 compositions required a cluster allocation.
  • Now: Screening 100,000 compositions fits in an afternoon.

Accuracy caught up

A decade ago, machine learning potentials were interesting curiosities. Their accuracy was OK for narrow chemistries they were trained on, and they were notorious for failing on out-of-distribution configurations. Universal foundation potentials changed that.

A model like MACE-MP-0, trained on millions of DFT calculations across the Materials Project, now achieves mean absolute errors on formation energies and forces that are within the inherent error of the underlying DFT itself. Equivariant architectures bake in the symmetries of physics, so they learn more from less data. And foundation checkpoints mean you do not have to train from scratch for every new system.

For a hands-on introduction, see our MACE-MP-0 guide.

Where MLIPs now win

There is a fairly well-defined set of workflows where MLIPs are simply the better tool in 2026:

  • Molecular dynamics. Long trajectories, large systems, realistic temperatures. MLIPs are 100,000 times faster than DFT MD with similar accuracy.
  • Structure optimization. Geometry relaxation of large systems, surface reconstructions, defect formation. A few hundred iterations finish in seconds instead of hours.
  • Property screening. Elastic constants, phonons, formation energies across a chemical space. MLIPs make sweep-based discovery practical.
  • Diffusion and rare events. Microsecond trajectories are now accessible, which means you can actually sample activation barriers.
  • Exploratory work on a laptop. You can open ASE, load a MACE checkpoint, and run a simulation on your MacBook at a coffee shop. See our laptop MD tutorial.

Where DFT is still essential

DFT is not going away. It is still the best tool for:

  • Reaction transition states. If you are computing activation energies for bond-breaking chemistry, DFT gives you the right answer while an MLIP may fail if the transition state is out of its training distribution.
  • Exotic electronic structure. Unusual spin states, strongly correlated systems, transition metal complexes with multiple low-lying configurations. Most MLIPs are not trained for these, and some DFT functionals themselves struggle.
  • Generating training data. The training data that MLIPs learn from comes from DFT. As long as MLIPs improve, DFT remains the reference standard that feeds them.
  • Novel chemistry. Work on elements, compositions, or geometries that are not represented in public training data. DFT is the safe default while you assemble a small reference set for later MLIP fine-tuning.
Warning
The biggest risk in using an MLIP for production is silent failure. An MLIP can produce beautifully reasonable numbers that are quantitatively wrong if the configuration is out of its training distribution. Always spot-check a subset of configurations with DFT, and prefer MLIPs that expose uncertainty estimates.

The hybrid workflow

The new best practice is hybrid. You use DFT to generate a small reference dataset for your chemistry, use that to fine-tune a universal MLIP, and then run production with the MLIP while periodically validating against DFT on novel configurations. This gives you DFT-quality answers at MLIP speed with controlled uncertainty.

A typical hybrid campaign looks like:

  • Run 200 DFT single-point calculations on diverse configurations of your system.
  • Fine-tune a MACE or NequIP foundation checkpoint on that dataset.
  • Run production MD or optimization with the fine-tuned model.
  • Spot-check 10 percent of accepted frames with DFT and check the residual error stays within tolerance.
  • If the error grows, add the offending configurations to the training set and retrain.

What this means for your group

If you are running a computational materials group in 2026 and every project still uses pure DFT for MD, you are leaving significant productivity on the table. Even a few weeks of effort to set up an MLIP-based workflow will typically pay for itself within one project by reducing turnaround time and increasing statistical sampling.

That does not mean everyone needs to become an ML expert. Hosted tools like the Materials studio on SciRouter let you call MACE-MP-0 and related universal potentials through a single API, without installing training frameworks or maintaining your own GPU infrastructure.

Bottom line

Machine learning potentials have crossed the threshold from interesting research tool to production default for most materials MD and optimization workflows. DFT is still essential as a reference standard and for the chemistry where MLIPs cannot be trusted, but the division of labor has fundamentally shifted. If you are planning a new project in 2026, the right question is not “which DFT functional should I use,” it is “which MLIP should I train or fine-tune.”

Explore the Materials studio →

Frequently Asked Questions

Are machine learning potentials actually replacing DFT?

For many practical workflows, yes. Production molecular dynamics, structure optimization, and property screening have largely moved to machine learning potentials in 2026 because they are a million times faster than DFT with accuracy that is within chemical tolerance for most systems. DFT is still the reference standard for training data and the right choice for reaction transition states, exotic electronic structure, and chemistry far from existing datasets.

How much faster are MLIPs than DFT?

Roughly a million times faster per single-point energy calculation, depending on system size and DFT settings. A DFT calculation that takes 10 minutes on a node can take 1 millisecond with an MLIP on the same node. This speed difference is what makes long-timescale molecular dynamics possible with ab-initio quality energies.

Are MLIPs as accurate as DFT?

On configurations that look like their training data, yes, typically within the intrinsic error of DFT itself. Far from the training distribution, MLIPs can fail silently. The safe pattern is to generate training data with DFT, train or fine-tune an MLIP on it, and run production on the MLIP while spot-checking against DFT on novel configurations.

What is the biggest risk in using an MLIP for production work?

Silent failure. A classical potential that is wrong usually gives obviously unphysical results. An MLIP that is wrong because the configuration is out of distribution can look completely reasonable while being quantitatively wrong. Always include uncertainty estimation or spot-check with DFT.

Which MLIP should I use?

Start with MACE-MP-0 as a universal baseline. If you need more accuracy for a specific system, fine-tune on DFT data for your chemistry. If you need massive scale, evaluate Allegro. If you need a foundation alternative, try Orb-v3. See our comparison post for a full breakdown.

Will DFT ever be obsolete?

No. DFT is the tool that generates the training data that MLIPs learn from. Even in a world where 99 percent of production runs use MLIPs, DFT is still the reference standard. The two tools are complementary, not competitive.

Try this yourself

500 free credits. No credit card required.