Supplementary Figures

Supplementary Figure 1 | TA data surfaces of TQ1:PC 71 BM for the indicated pump-fluences. Solid lines indicate the bleach-signal-maximum, which is assigned to the thermalization of the holes in the disorder broadened density of states 1,2 , see Supplementary Fig. 2. The formation of charge-transfer states and free charge carriers in TQ1:PC 71 BM occurs on a time scale of hundreds of fs or less as can be inspected from the photo-induced absorption band. The barely visible boundary line at ≈3 ns is due to a change from a mechanically to an electronically delayed excitation pulse, scaling factors are indicated in the figure.

Simulations are nominally identical to those described in the main text and Supplementary Fig. 7, except for the applied electric field, which is now harmonically modulated (AC) -a sinusoidal electric field E(t)=E 0 sin(2πft) at the indicated frequencies f was applied. The amplitude of the AC electric field, E 0 , was chosen to match that encountered in THz experiments ≈10 5 V/m 8 . Simulated mean mobility is shown in the figure, mostly dominated by the higher electron mobility. Note, that at frequencies used in TRMC experiments ≈10 10 Hz AC effects are unimportant, as the simulated mobility trace at 10 10 Hz overlaps with the DC simulation and the TRMC/TA experiment. Mobility is higher at higher frequencies as shorter distances are probed 9-11 : at high frequencies carriers predominantly hop backand-forth between sites of similar energy, and are thus less prone to encounter high energy barriers (difficult hops), as is the case for long-distance DC transport probed by TRMC/TA and pCELIV. However, full treatment of non-equilibrium AC conductivity is beyond the scope of this work. reached thermal equilibrium during their presence in the device, for example compare the electron DOOS to their estimated equilibrium energy σ 2 /kT, indicated by the dotted blue line. Also the hole population in PCDTBT:PC 61 BM has not reached its equilibrium energy σ 2 /kT due to the relatively large disorder of the PCDTBT HOMO. Although the above figure indicates that in case of TQ1:PC 71 BM the hole population has reached its equilibrium energy σ 2 /kT (dotted red line), it must be noted that 1) The relatively slow transition to this state is not visible in the figure, however charge carriers still have to undergo thermalization to reach this state, Figure 1a of the main text shows this transition.

Supplementary
2) The lesser thermalized charge carriers generate most of the photocurrent via their considerably higher (time-dependent) mobility and are thus extracted before reaching thermal equilibrium, see the corresponding energy level-resolved photocurrent in the right plot. This is generally the case, for example compare for other populations.
3) The experiment of Figure 1a indicates further thermalization that is not captured in the model prediction, see Figure 1a.

Supplementary Tables
Supplementary Table 1 | Overview of TQ1:PC 71 BM parameters used in the simulations. Parameters in grey have no effect on the results but are shown for completeness. HOMO and LUMO refer to the orbitals of the effective medium.

Parameter
Value + unit Experiment We use the Miller-Abrahams expression to quantify, with the least number of parameters, the nearest-neighbor hopping rate of a charge carrier from an initial state i with energy to a final state f with energy as Here is the external electric field, the vector connecting initial and final sites, the attempt-to-hop frequency, and q the positive elementary charge. The + (−) sign refers to electron (hole) hopping. The term is the change in Coulomb energy and is calculated by explicit evaluation of the interaction of the moving charge with (a) all other charges in the simulated device and (b) their image charges, as well as of the interaction of the image charges of the moving particle with (c) the particle itself and (d) all other particles. Image charges arise when metallic contacts are present; the number of image charges accounted for in the simulations is increased till the resulting effective Coulomb potential no longer changes. In order to avoid divergences at zero separation, the Coulomb interaction between a pair of (unlike) charges ⁄ with the dielectric constant ( = 3.6) and the electron-hole distance is truncated at minus the approximate exciton binding energy of = 0.5 eV. The single-particle site energies are drawn from a Gaussian distribution function √ ( ) with the mean energy and the broadening of the total density of states (DOS) . The HOMO and LUMO energy of a single site are assumed to be uncorrelated.
We define an effective hopping medium with (different) electron and hole hopping parameters and that correspond to the donor HOMO and acceptor LUMO levels, respectively. The driving force for charge transfer is implemented via an on-site electron-hole repulsion with a magnitude that equals the LUMO level offset between donor and acceptor. More elaborate accounts of the bulk heterojunction double the number of unknown hopping parameters and force one to make statements about the hierarchical length scales in the phase separated morphology. More importantly, the transport and relaxation, in which we are mainly interested here, occurs in the levels that are explicitly accounted for in the present treatment.
A consequence of the neglect of the phase separated morphology is that the model would not be able to reproduce an eventual reduced bimolecular recombination rate due to electrons and holes being spatially separated in different phases. However, at the conditions investigated here 'nonreduced' bimolecular recombination is already unimportant.
Spatially direct excitons, formed by an electron and a hole on the same site, can recombine with rate . Similarly, when sitting on neighboring sites they form a CT complex that can recombine with rate . This implies that mono-and bimolecular recombination are treated on equal footing, as recombination rates of exciton and CT species do not depend on the history of the constituent charges. Exciton diffusion by the Förster resonant energy transfer (FRET) mechanism is explicitly accounted for. The transition rate is evaluated as is the Förster radius, the radiative exciton decay rate, the Heaviside step function and , the exciton energies at the initial and final sites. Dextertype exciton diffusion is implicitly accounted for as a double charge hopping process.
The waiting time before an event (hop or recombination) occurs is calculated as where is a random number drawn from a homogeneous distribution between 0 and 1 and is the sum of the rates of all possible events. The event that occurs after is selected randomly, using the rates of all possible events as weight factors. Energies, rates, and waiting time are recalculated after each event.
In order to simulate the transient response to a short light pulse, each simulation is started by creating an initial concentration c 0 of excitons on an otherwise empty calculation grid. Given the high photon energy employed in the experiments an equal excitation probability is assumed for all sites, in other words excitons are generated at fully random positions. Both in pulsed and steady-state simulations all charges are tracked in time, energy and position from the moment of generation till extraction or recombination.
Periodic boundary conditions in the x,y-directions were applied for both charge motion and (image and direct) Coulomb interactions; contacts laying in the z-plane are included unless stated otherwise and were implemented as perfect sinks. Independence of the results on the box size used in the calculations was assured.
As MC simulations are too time-consuming to allow true least-square fitting, estimates of the parameters were made entirely on basis of visual inspection of simulated data and kept unchanged during further simulations. Table 1, 2 were set upfront as determined in previous work 2, 16,17 and kept fixed during all simulations. The lattice constant (= hopping distance) was fixed to a reasonable value that was previously extracted from curve fitting to SCLC data 6,15 . For PCDTBT:PC 61 BM the lattice constant was further adjusted to better match the charge carrier recombination dynamics 2 . For TQ1:PC 71 BM the ratio of the Förster radius and the lattice constant turned out to have no noticeable effect on the simulations and was set to a value of 2.2, which implies a Förster radius of 4 nm. Recombination rate for excitons was set to 1.9x10 9 s -1 as determined from TRPL in Supplementary Fig. 3. The rate for CT complexes was set to 1x10 8 s -1 to reproduce the experimental IQE≈0.85 13 . For PCDTBT:PC 61 BM the Förster radius, the lattice constant and the recombination rates for excitons and CT complexes were adjusted to match the TRPL measurements ( Supplementary Fig. 3) and experiments on charge carrier recombination dynamics 2 . See Supplementary Table 1, 2 for which experiments defined the simulation parameters.

Most of the parameters in Supplementary
As is generally understood as related to some phonon frequency, it is not upfront clear why the different geometries of fullerene and polymer (but with somewhat similar carbon-carbon bonds) should give so different attempt frequencies. It should however be kept in mind that in the highly simplified Miller-Abrahams formalism (Eq. 1) also the overlap or transfer integral is contained in  0 . It does not seem unlikely that the overlap integral of the rate limiting hops differs substantially between the spherical PCBM and the highly anisotropic polymer. Similar differences in as obtained here were found from a detailed analysis of SCLC curves of pristine PC 61 BM and pristine MDMO-PPV 17 .
Although our model is a crude approximation of ultrafast (and in literature suggested to be coherent) phenomena occurring on the fs-ps time scale, within the framework of our model the used parameters predict charge-transfer rates that are comparable to the experiment of Gélinas et al. 20 . As our lattice is 6-fold coordinated, within the framework of our model charge-transfer occurs approximately after a time 1/(6 x v 0 ), where v 0 is the attempt-to-hop frequency. Thus, following photo-generation, charge transfer occurs in ≈17 fs in TQ1:PC 71 BM and in ≈56 fs in PCDTBT:PC 61 BM. These are reasonable values as compared to the measured ≈40 fs in PCDTBT:PC 61 BM (Gélinas et al. 20 ). Also, detailed treatment of ultra-fast phenomena is not the main purpose of our model. The purpose is to use a single framework to describe a large number of experiments over a wide range of times, and to accomplish this by the least number of parameters -most of which are in fact determined by experiments.
Finally, we point out a technical detail when comparing the simulation parameters listed in Supplementary Table 1, 2 to those from other theoretical models. Specifically, the disorder value and attempt-to-hop frequency v 0 that are used to successfully fit the experimental data employing the above model will differ from the disorder value that might (or might not) fit the experimental data using non-adiabatic Marcus theory for charge or exciton dynamics.

Supplementary Note 2 | Comparison of PA and PB bands for measuring hole thermalization.
We point out that the photo-induced bleach (PB) band is the more appropriate choice to monitor hole thermalization dynamics as compared to the photo-induced absorption (PA) band. We consider the PA band to be more problematic due to the likely overlap of several contributing species: singlet excitons, triplet excitons, polarons, possibly not only from the donor but also (though weaker) from the acceptor. In recent literature the relative contributions of these species are disentangled via multivariate curve resolution (MCR) methods. However, such methods cannot treat dynamic spectra: the main assumption of MCR analysis are static spectra in energy of each individual species, which is not suitable for the present purposes. Moreover, the PA band probes the absorption between excited states, the absorption of which may (or may not) change as the carrier population thermalizes, making it difficult to quantify the energy loss to thermalization. Via the PB signature we are predominantly monitoring the energetic position of the hole, making the PB band the appropriate choice in this case.

Supplementary Note 3 | Explanation of how the TRMC/TA mobility was estimated.
Since both of the THz and TRMC techniques in fact monitor the time-resolved conductivity σ(t) = en(t)µ(t), for the correct determination of the time-dependent mobility µ(t) the charge carrier concentration dynamics n(t) in the same time range must be known. This was accomplished by combination with the timeresolved photo-induced absorption (PIA) data 5 , allowing for the time-dependent mobility µ(t) in the commonly accessible time range to be estimated. As PIA only follows the hole population dynamics such analysis was made possible by the following observation: for TQ1:PC 71 BM a purely bimolecular charge carrier concentration decay is observed 5 -charge carriers recombine in pairs, thereby both hole and electron densities should be similar. This allows for the direct conversion of the timeresolved conductivity σ(t) to the time-dependent mobility µ(t) by the use of the PIA experimental estimate of n(t). Similar arguments were used in case of THz, see Supplementary Fig. 8.

Supplementary Note 4 | Explanation for the high number of hops.
The high number of hops during the diffusion-dominated loss step is because most of the hops occur downward in energy. This is confirmed by inspecting the minimum time it would take for an average charge carrier to complete such a number of hops, as can be done by the use of simulation parameters in Supplementary Table 1, 2. For example in case of electrons in TQ1:PC 71 BM the number of hops (≈7000) times the inverse of the attempt to hop frequency ( ⁄ , an estimate of the minimum time it takes between consecutive hops) results in 7000 × ⁄ ≈0.7 ns. This is in excellent agreement with the transition from diffusion-to drift-dominated motion as can be read from the thickened simulation traces in Figure 2 of the main text. The same analysis leads to 0.36 µs and 80 ns for the holes in TQ1:PC 71 BM and PCDTBT:PC 61 BM respectively, again in good agreement with Figure 2 of the main text. Thus, during the diffusion-dominated loss step, an average charge carrier mostly hops downward in energy.