Thermostats and barostats (system regulators)¶
This page illustrates characteristics feature of various types of regulators used in molecular simulations to maintain the temperature and pressure of the system.
The text will make references to DL_POLY but it is also applicable to other packages.
There is a range of integration techniques available in DL_POLY and other MD packages to bring and maintain a model system on N particles to a targeted volume (V), energy (E) or temperature (T) via NVE (microcanonical) and NVT (canonical) ensembles, and pressure (P) via NPT and NsT (isobaric-isothermal) ensembles.
Thermostats help drive the model system to the desired target temperature, \(T_{target}\), and need a user-specified relaxation time or a friction frequency, depending on its definition within the equations of motion. Barostats help drive the system to the desired target pressure, \(P_{target}\), and also need a user-specified relaxation time or a friction frequency, depending on its definition within the equations of motion. These time-related constants regulate the eagerness with which the thermostat and/or barostat (collectively called the ‘system regulators’) apply control onto the system in order to maintain its temperature and/or pressure (collectively called the ‘system parameters’) to the desired targeted values (thermodynamic state). In general, the larger the value of the relaxation time (the smaller the frction frequency), the slower the response of the regulators to maintain the targeted parameters.
If the relaxation time is too large then the integrator will allow the system to relax but it will be difficult to maintain the desired system parameters close to their targeted values. This could lead to the system responding more slowly to fluctuations in the system parameters or obtaining values offset from the targeted values.
If the relaxation time is too small the integrator will maintain the system parameter too vigorously and it may not allow the system to relax properly or even lead to large uncontrolled fluctuations of the system regulators around their targeted values.
Note
There are no unique ‘ideal values’ for the relaxation times: their choice is dependent on the type of system being modelled and the nature of the regulators (equation of motions). However, for many atomistic systems a typical value for a thermostat relaxation time is within the interval of 0.2 to 2.0 ps (e.g. 0.75 ps) and a typical value for a barostat relaxation time is within the interval of 0.5 to 4.0 ps. (e.g. 1.5 ps).
Check the temperature of the system in the STATIS or OUTPUT file to see if you are happy with the temperature and pressure fluctuations.
It is worth noting that not all integrators generate correct ensembles and hence not all of them can sample correctly the equilibrium or be used correctly for non-equilibrium simulations. However, all of them have their own uses and it is generally a matter of preference which one should be used, especially for equilibration.
Temperature rescale¶
Another way to maintain the temperature of the system is by using brute-force rescaling of the particle velocities to maintain precisely the specified target temperature (given in the CONTROL file) at user-specified intervals of time. This option is only recommended for equilibrating the system at the initial stage. Since this option interferes with the integrators’ temperature fluctuations, the forced dynamics do not correspond to any ensemble dynamics and any properties measured while using this option should not be used for system sampling.
Temperature rescaling will be activated in DL_POLY by using the directive equilibration. Such rescaling only applies while the MD timestep number is less than the specified number of equilibration steps in the CONTROL file. Once the timestep number exceeds the number of equilibration steps (i.e. when the number of steps is larger), only then are the system regulators from the selected integrator applied.
Note
Example below shows an example how to apply brute-force temperature rescale effect in the DL_POLY CONTROL file:
steps 80000
equilibration steps 50000
This means brute-force scaling temperature applies for the first 50,000 steps. After that, the selected system regulators apply for the remaining 30,000 steps.
Types of system regulators¶
Different types of system regulators have different characteristics that perform suitably depending on the system types. There is a vague consensus, often with weak scientific argument but largely an empirical ‘community-led’ preference, to choose one type over the others. The preferred type for equilibration purposes is the Berendsen thermostat as it controls each particle velocity directly down to target temperature. This could lead to ‘the flying icecube’ effect if no control is exercised to the centre of mass motion gathered by this algorithm. However, this undesired effect is counteracted by default in DL_POLY.
Two broad types of system regulators are deterministic and stochastic. From a given initial state, deterministic regulators act consistently (at least to machine precision) on a system regardless of how a simulation is run.
Stochastic regulators depend upon the use of pseudorandom number generators, whose setup may affect the results of a simulation. That said, (i) generators and their initial states can be selected to give consistent results for different runs, and (ii) regulators with correctly defined dynamics should give statistically similar results even for runs with different generator setups.
Coupling methods¶
Evans thermostat¶
It couples the particles in the system to a heat ‘bath’ via a kinetic temperature constraint that generates a self-adjusting velocity (kinetic) friction. It generates a proper NVEkin ensemble and can be used for systems with a steady flow. It can be used for equilibration. It is suitable for soft condensed matter systems such as liquids.
Langevin thermostat¶
It couples the system to a coupled viscous background (with a user-defined kinetic friction) and a stochastic heat ‘bath’ (with random frictional forces). The effect of the algorithm is to thermostat that system on a local scale – ‘cold’ particles are given energy whereas ‘hot’ particles are cooled down. It generates a proper NVT ensemble. It is suitable for soft condensed matter systems such as fluids and such with varying density. The thermostat is suitable for biological systems (e.g. proteins and water) as well as solids in condensed phase. It can use larger time steps than other thermostats. The thermostat adds a random centre of mass motion that has zero mean, i.e. there will not be a long-term drift of the centre of mass.
The damping effects apply on the particles means there is a lost of momentum transfer. Therefore, diffusion coefficients cannot be determined when applying the Langevin thermostat.
Andersen thermostat¶
It randomly selects particles and replaces their velocities with values randomly chosen from a Maxwell-Boltzmann distribution for the target temperature. Stochastic collision operator determines the number of particles whose velocities are replaced. The thermostat is not a true NVT ensemble. However, it is still suitable and popular for soft condensed matter systems such as fluids and biological, provide it is not used for the calculation of exact dynamical properties. If no care is taken the original algorithm may lead to centre of mass drift. However, this is counteracted in all modern versions of the algorithm.
Berendsen thermostat¶
It couples the system to a virtual heat ‘bath’ via a brute force velocity rescaling, only moderated by the relaxation time parameter. Suitable for equilibration. However, its direct rescaling of velocities results in wrong equipartition of energy components that leads to systematic global motion of the system (flying ice cube). This effect is counteracted nowadays but it does not sample correctly the NVT ensemble and as such is non-ergodic. In general, it is thought the Berendsen approach is quite suitable for equilibration purposes (e.g. a steep descent to \(T_{target}\)) but not recommended for sampling. Although, if a system is well defined by a model and in equilibrium, all integrators should generate the same energies and virials.
Nosé-Hoover thermostat¶
It defines a mass for the heat ‘bath’ governed by the relaxation parameter. It results in slow, decreasing fluctuations of the system kinetic energy towards the desired \(T_{target}\). It samples the NVT (canonical ensemble) and is considered the gold standard by many communities. It is suitable for all condensed matter systems (liquids, solutions, biological systems).
GST thermostat¶
The Gentle Stochastic Thermostat (GST) is an extension on the Nosé-Hoover thermostat in which the Nosé-Hoover thermostat friction has its own Brownian dynamics, governed by a Langevin friction. The GST samples the NVT (canonical ensemble) and is considered the new gold standard. It is the only ensemble that warrants ergodicity by construction and thus is suitable for all model systems.
DPD thermostat¶
Dissipative Particle Dynamics (DPD) is a pairwise generalised form of the Langevin thermostat. It applies random (‘heat bath’) and frictional (viscous) forces to pairs of particles within a cutoff. It conserves both system-wide and local momentum and is thus Galilean invariant, i.e. it does not produce global motion (flying ice cube) for static systems while providing the correct hydrodynamic behaviour if flow fields are applied. Dissipative (friction) force terms can be related to fluid viscosity and self-diffusivity, although relationships are complex and affected by particle interactions. It can use larger time steps than other thermostats. It is suitable for soft condensed matter systems (including those with flow fields) at coarse grain length scales, intended for mesoscopic (DPD) simulations.
Note
To obtain equilibration as quickly as possible in DPD simulations, the dissipative force parameter \(\gamma\) is often set to \(4.5 \frac{k_B T \tau}{r_c^3}\) (where \(k_B T\), \(\tau\) and \(r_c\) are the selected energy, time and length scales for the simulation). This value corresponds to a minimum fluid viscosity and maximum diffusivity.
Berendsen barostat¶
It couples the system to a virtual external pressure ‘bath’ by rescaling the simulation volume and particle positions according to how far the instantaneous pressure is from the target value \(P_{target}\). The scaling factor is related to the ratio of the system isothermal compressibility and the barostat relaxation time parameter. DL_POLY uses the isothermal compressibility of liquid water as a representative value for the barostat. The Berendsen barostat is often used along with the Berendsen thermostat (as in DL_POLY) to generate constant pressure and temperature (e.g. NPT) ensembles, although it can be coupled with other thermostats (e.g. with DPD in DL_MESO). In general, it is thought the Berendsen approach is quite suitable for equilibration purposes (e.g. a steep ascent or descent to \(P_{target}\)) but not recommended for sampling.
Langevin barostat¶
It couples the system to a virtual piston, which moves according to how far instantaneous pressure is from the target value \(P_{target}\). The piston is also coupled to a viscous background with a user-defined kinetic friction and a stochastic heat ‘bath’ with random frictional forces. The piston mass and the kinetic friction are both determined from a barostat relaxation time parameter, while the random frictional forces are scaled according to the barostat kinetic friction and target system temperature \(T_{target}\). The Langevin barostat is often used along with the Langevin thermostat (as in DL_POLY) to generate constant pressure and temperature (e.g. NPT) ensembles. It can frequently obtain the target pressure more quickly than other barostats and reduce correlation times for sampling of system properties.
Note
Care needs to be taken when selecting the Langevin barostat relaxation time parameter, particularly for the choice of piston mass. Overly large values can decouple the barostat and particle dynamics (leading to inefficient sampling), while too small values can disturb the particle dynamics. The optimum value normally has to be chosen by trial-and-error, although a rule of thumb suggests the resulting frequency of volume fluctuations should be around ten times smaller than the thermostat frequency (related to its kinetic friction parameter). In turn, a good choice for barostat kinetic friction is between half and one tenth of the volume fluctuation frequency.
Nosé-Hoover barostat¶
It couples the system to a virtual external pressure ‘bath’, whose mass is governed by the relaxation parameter. It rescales simulation volume and particle positions, using how far the instantaneous pressure is from the target value \(P_{target}\) to evolve the rescaling factor. Coupled with the Nosé-Hoover thermostat, it samples NPT and other isobaric-isothermal ensembles. It is suitable for all condensed matter systems (liquids, solutions, biological systems), although it only acts correctly for larger simulations. An extended form of this barostat with further connections to the thermostat – the Martyna-Tuckerman-Klein barostat – overcomes this restriction.