A useful way to understand AlphaFold 3’s sampling behavior is to look not only at the final predicted structure, but at what happens along the reverse diffusion trajectory itself. If we track quantities such as the physical energy of samples, noise scale, and update magnitude over time, a very clear pattern emerges: structures remain physically imperfect for most of sampling, and only take proper global shape in the final low-noise steps.

This behavior is a result of the diffusion procedure implemented in Algorithm 18, Sample Diffusion, which follows an EDM-style sampler with churn. Rather than simply marching monotonically from noise to structure, the sampler repeatedly perturbs the current coordinates, denoises them, and then takes a Euler-like update step. Because of the churn mechanism, AlphaFold 3 deliberately injects additional noise during part of the trajectory, which encourages exploration but also delays local geometric convergence. This mechanism is shown in step 4 -7 of the Sample Diffusion Algorithm from Alphafold3 Supplementary Information.

To better understand how samples evolve under the native AlphaFold 3 denoiser, we can track four quantities throughout the trajectory:

(a) Constraint Energy
This plot shows the total physical energy of the samples over the reverse diffusion process. It stays fairly high and noisy for much of the trajectory, then drops sharply near the end. This means physically invalid geometry is tolerated for a long time before being cleaned up in a late refinement phase.

(b) Number of Violations
This tracks counts of clashes, bond violations, angle violations, and related physical defects. Like the energy plot, the counts remain elevated for most of sampling and then collapse late. So AlphaFold 3 does not fix violations steadily; it fixes many of them only near the end.

(c) Noise level (t)
This shows the effective diffusion noise scale decreasing over time. Early on, t is very large, so coordinates are still heavily corrupted and the model is operating in a coarse denoising regime. Around step 160, the trajectory reaches approximately t = ~ 1. After that, it enters the t < 1 regime, where the residual noise is small enough that the model can focus on fine local corrections instead of just global organization.

(d) Drift norm
This measures the size of the denoiser-driven update. It stays relatively flat for much of the trajectory, then drops abruptly near the same point where t becomes small. That suggests the denoised prediction becomes much more self-consistent late in sampling. The slight rebound at the very end is expected, because dividing by a very small t_hat can make even tiny residual mismatches look larger.

What happens when t reaches 1 around step 160

This is the key transition.

When t > 1, the sample is still noisy enough that AlphaFold 3 is mainly handling global structure: chain placement, fold organization, interfaces, and ligand positioning. In this regime, local physical problems such as clashes or poor stereochemistry can persist because the model has not yet committed to a precise geometry.

When t = ~1 and then moves below it, the sampler enters a fine refinement regime. Now the remaining perturbation is small relative to the coordinate scale of the structure, so the model can start making precise local corrections. That is why, right after about step 160, both the constraint energy and the number of violations collapse rapidly.

So the main takeaway is:

AlphaFold 3 builds the global structure first, and only cleans up local physical validity once the noise scale becomes very small.