Mark Tuckerman talk: Free energy based structure prediction

Mark Tuckerman (MET) visited London on the 31st March 2016, and gave a highlight seminar to the TYC at UCL. I had been really enjoying his recent (2010) Computational Stat Mech book [1]. I know very little about structure prediction, but it is of general and captivating interest, and an essential step in truly predictive computational materials science. Here are a summary of my notes from his talk, written up mainly for my own reference, and for clickable citations.

A point Mark made at the start and repeatedly was that we need to go beyond the static structure picture - we need the entropic contribution, and we therefore need to be able to calculate free energies.

Small molecular crystals

The first area discussed was predicting structure of small molecular crystals, the example for them was having taken part in the CCDC Blind Test [2], but focusing on a single molecule & attempting to explore the polymorphs. The motivation here is that for Pharma research, knowing the existence of energetically accessible polymorphs is of utmost importance due to highly varying bioavailability of different polymorphs.

The method started off fairly standard. As general force fields are too coarse, and quantum mechanical calculations too expensive, a customised force field was developed. QC calculations on dimers were used to parameterise, with a fairly normal PBE0+TS; 6-311g**(d,p). 25000 Unit cells were generated with UPACK, a software package I had not heard of, but which seems pretty extensive [3].

MD (I think just vanilla) on the packed structures lead to trajectories where interconversion from one structure to another could be seen. Eventually they settled down to 50 (lowest energy?) structures, plotted on an energy vs. density plot.
Their 4th lowest energy structure agreed with the ‘blind’ experimental structure. Mark was more pleased with having found a lot of low lying polymorphs.

So far so good for an interesting seminar, and a laudable result - but this is where it got really interesting.

Mark linked from structure prediction to protein folding via a 2004 PNAS [4], which proposed that both problems are essentially the same. You have a rough energy landscape, you need to determine the occupancy of the basins.

Collective variables

To do this, Mark suggests targeting collective variables - something that distinguishes different conformal states. These are selected ‘by hand’ with ‘insight’ and ‘chemical intuition’ to fit the problem studied.

  • The most simple example are the Psi and Theta angles of the Ramachandran plot [5] for amino acid residues.
  • For crystals you can take the unit cell vectors and stuff them into what Mark referred to as the h matrix.
  • For packing, you can take angles and distances, or use Q, W Steinhardt bond-order parameters.
    • [Incidentally, I recently learned thanks to another seminar that these seemingly mystical parameters are really quite simple - they’re just spherical harmonics! You consider only even-l versions of these, making them invariant under inversion. [6] ]

Temperature accelerated MD

Temperature accelerated MD was then introduced. Mark described it as essentially being an additional (standard) equation of motion being simultaneously solved, in the collective variables. A very large fictitious mass is used for the collective-variable centre, and a separate thermostat (~32000 K was an example). This physically means that the collective variables have the energy to hop potential barriers between minima, and the momentum to smoothly move between them. The large difference in mass allows for an adiabatic decoupling of the Equations of Motion.

I did not realise that temperature accelerated MD was this conceptually simple.

Metadynamics is an alternative way to explore the free energy landscape, but relies on the simulation dropping little Gaussians in the potential energy where they are currently, so requires a certain amount of time to fill up the basins and escape to the next one. To get the free energy, once you’ve filled all the basins you invert the Gaussians and this gives you the free energy surface.

Protein folding: Alanine Decamers between alpha helix and beta hairpin

Accelerated MD on this as a model system (I think with the CHARMM forcefield), with the Ramachandran variables (20 collective variables). Running this produces lots of structures, with the transition paths identified [10]. Using a Berendsen thermostat with added noise [7], defining a path as the collective variable, you can integrate along to get free energy. This gives you a smooth energy along the coordinate, and thus a well found value for the free energy difference.

I found a 2014 YouTube video containing the MD, but the extraction of free energy seems a bit different:

Polymorphic benzene

Separately they studied polymorphs of benzene. The ‘h’ matrix (lattice basis) was explicitly evolved in the EoM, dragging atoms instantly with it. With benzene at 100K (atomic thermostat) and 32000 K (collective variable thermostat), 5 ns of MD was sufficient for all free energies, 0.5 ns of MD found all the polymorphs.

This is explored in [8], and MET references thereafter.

Melting Copper

Melting Cu, with collective variables on the diagonal of h; and a Q4, Q6 spherical harmonic ‘bond order’. The EAM potential was used.

Here they looked at classical nuclear theory. The simulations showed a more complex story - the hot solid had defects which were highly mobile. Defects formed clusters to minimise their free energy

You can’t visualise the surface - it is too high in dimension. But you want to see the ‘interesting’ minima and saddle points. You can use a Parrinello-group sketchmap [9], which appears to be a clustering graph representation. Here [11] they used these ideas, and the START method where the free energy gradient and the Hessian were constructed in terms of collective variables, to give you the weakest eigenvector to walk along with steepest descents.

Some very recent and unpublished data investigated Anthracene crystals, defining the Euler angles (though a later figure mentioned an ‘absolute quaternion’) and cell parameters as collective variables. They found the experimental structure, and loads of near-energy polymorphs


It was an extremely thought provoking talk. The explanation of temperature accelerated molecular dynamics was the first time that it made sense to me. The power of integrating the EoM for a collective variable & so ride roughshod across a rough potential energy landscape with the ability to extrapolate free energies is clearly fantastically powerful.

One bit of black magic that I do not understand, is where the force acting on the collective variables comes from. Some kind of potential defined in the collective variables? By direct coupling between the atomistic positions and the collective variables?

However, I felt that everything presented seemed a little clunky. Certainly the science was fantastic, but every example seemed to have been set up slightly differently and individually by the researchers. There was no mention of these codes or methods being made available.

One thing obviously missing from a scientific point of view was a algorithmic way to define the collective variables, rather than relying on intuition.

In a paper I’ve been reading recently [12], the authors of VoroTop introduce a method to identify local structure independent of radial distribution functions and bond orders. Instead they define the 3D Voronoi cell for individual atoms, and then compare the isomorphism (or not!) of the 2D Schlegel diagram (i.e. a planar graph of connectivity of the Voronoi vertices) to identify topologically distinct local bonding. I wonder if some ideas from here, could be borrowed and used to construct collective variables automatically.

Once you can identify the collective variables automatically, the exploration of rough potential energy surfaces, while retaining the free-energy of structures, will be practically pleasant.

With a bit of Googling I found a Tuckerman slide deck from 2014 which overlapped with this talk a fair bit:

Mark has a Google Scholar page, for if you want to dig through more of his literature [13].


[1] Mark E. Tuckerman (2010), Statistical Mechanics: Theory and Molecular Simulation. Oxford Graduate Texts. (



[4] “Exercises in prognostication: Crystal structures and protein folding”. Jack D. Dunitz and Harold A. Scheraga (2004). PNAS 40(101), 14309-14311.


[6] “Bond-orientational order in liquids and glasses” Paul J. Steinhardt, David R. Nelson, and Marco Ronchetti (1983). Phys. Rev. B 28, 784.

[7] “Canonical sampling through velocity rescaling”. Giovanni Bussi, Davide Donadio and Michele Parrinello (2007). J. Chem. Phys. 126, 014101 (2007);

[8] “Temperature-Accelerated Method for Exploring Polymorphism in Molecular Crystals Based on Free Energy” Tang-Qing Yu and Mark E. Tuckerman (2011). Phys. Rev. Lett. 107, 015701;

[9] “Simplifying the representation of complex free-energy landscapes using sketch-map”. Michele Ceriotti, Gareth A. Tribello, and Michele Parrinello (2011). PNAS 32(108), 13023-13028.

[10] “Sampling saddle points on a free energy surface” Amit Samanta, Ming Chen, Tang-Qing Yu, Mark Tuckerman and Weinan J. Chem. Phys. 140, 164109 (2014);

[11] “Locating landmarks on high-dimensional free energy surfaces” Ming Chen, Tang-Qing Yu, and Mark E. Tuckerman (2015). PNAS vol. 112 no. 11 (2015);

[12] “Topological framework for local structure analysis in condensed matter” Emanuel A. Lazar, Jian Han, and David J. Srolovitz (2015). PNAS vol. 112 no. 43;

[13] Mark E. Tuckerman, Google Scholar.

Jarvist Moore Frost
Electronic structure theory