To make these measurements, we create an atomic realization of the three-dimensional (3D) disordered Bose–Hubbard model (DBHM) using ultracold 87Rb atoms trapped in a disordered optical lattice9. The DBHM has been used to describe systems as diverse as 4He in disordered media, Josephson-junction arrays, and long-wavelength properties of superconducting electron pairs. In the DBHM, a strongly interacting superfluid (SF) undergoes a quantum phase transition (QPT) into a Bose glass (BG) when subjected to disorder. The critical exponents of this transition fall in the percolation class10,11. The BG phase exhibits the peculiar property of lacking long-range order while possessing infinite superfluid susceptibility1, and it is therefore viewed as a gapless insulator with finite compressibility that arises from the presence of quasi-condensates, or SF puddles, embedded in an insulating background. Disordered ultracold atom gases have been used to indirectly measure the SF–BG transition by means of transport and coherence measurements in 1D (ref. 12) and 3D (ref. 13) disordered lattices and in 1D quasi-periodic lattices14,15.

In our experiment, we prepare a gas consisting of (27 ± 2) × 103 atoms cooled below the critical temperature Tc for Bose–Einstein condensation in a parabolic trap such that the condensate fraction is greater than 90%. A disordered cubic optical lattice formed from pairs of counter-propagating λ = 812 nm laser beams and a 532 nm optical speckle field is superimposed on the gas (see Supplementary Methods). The atoms experience a potential energy shift proportional to the speckle intensity, which varies randomly in space, leading to disorder in the Hubbard parameters. The DBHM we realize is characterized by the Hamiltonian

where i and j index the lattice sites, 〈〉 indicates that tunnelling occurs only between adjacent sites, and tij, Ui and εi are the Hubbard energies. In equation (1), is the number of particles on site i, () removes (adds) a particle from site i, m is the atomic mass, ω is the geometric mean of the trap frequencies, ri is the distance to the centre of the trap, and μ is the chemical potential. We measure all energies in terms of the recoil energy ER = h2/22kB × 170 nK. The distribution of the Hubbard parameters, which are broadened around the values for the uniform system, are precisely known9,16. The strength of the disorder is characterized by the average potential energy Δ associated with the speckle, which is approximately equal to the standard deviation of the εi distribution. The lattice potential depth s (which controls U/t) and Δ are independently adjusted by tuning the power of the lattice laser and 532 nm light. For the values of s sampled in this work, the gas is a strongly correlated, quantum-depleted SF when Δ = 0. We do not explore sufficiently high s to generate a Mott insulator phase in the gas.

We probe the BG–SF transition by measuring the amount of excitation produced by quenching Δ at fixed s. The disorder strength is linearly ramped from an initial value of Δ0 to zero in 30 ms (Fig. 1b), which is slow enough to avoid creating excitations solely through the time variation of the spatially inhomogeneous disorder potential (see Supplementary Methods). On the basis of general arguments regarding the phase diagram in untrapped systems, the BG phase will appear in the low-density edge of the gas for sufficiently high Δ0 (ref. 1). For stronger disorder, the BG–SF boundary moves inwards, encompassing more of the atoms. Excitations produced by the quench are measured using time-of-flight (TOF) imaging. By imaging after a long (50 ms) period of free expansion, vortices and any other excitations involving velocity fields (including phonons) are transformed into modulations of the density profile and the measured optical depth (OD). These excitations are visible in the characteristic images shown in Figs 1c and 2b. For low Δ0, the density profile after the quench and TOF is smooth, whereas for high Δ0, features consistent with vortices are present.

Figure 1: Quench through BF–SF QPT.
figure 1

a, Schematic phase diagram of the DBHM and quench. The gas is quenched from the BG to the SF regime by rapidly reducing (green arrow) the disorder strength Δ from Δ0 to zero at fixed U/t, which is determined by the lattice potential depth s. Equilibrium configurations and the disordered lattice potential (false colour) are shown at three values of Δ. For sufficiently high Δ, BG (blue) and SF (light grey) phases coexist in the trap. b, Time sequence for the measurement. The lattice potential depth and disorder strength are shown using red and green lines. c, Equilibrium is disrupted during the quench and excitations are produced, which are measured in TOF images (greyscale). Images are shown for Δ0 ≈ 0.5ER (i) and Δ0 = 0 (ii) at s = 12ER. The white ellipse indicates the fitted Thomas–Fermi radius, and the fitted profile is subtracted in the lower images. For sufficiently high disorder, excitations such as vortices are apparent (red arrow) after the quench, whereas smooth profiles are obtained at low Δ0. Clear images of vortices are rare, because vortices generated by the quench are randomly oriented relative to the imaging direction.

Figure 2: Sample experimental data and numerical results.
figure 2

a,b, Results of quench measurements. a, The observed χ ̃2 as Δ0 is varied at s = 11ER and the piecewise-linear fit (red line) used to extract the threshold disorder Δth are shown. The error bars show the standard error in the mean for the 6–12 measurements averaged at each Δ0. b, TOF images obtained after the quench are shown for Δ0 = 0 (i), Δ0 = 0.7 (ii), and Δ0 = 0.95ER (iii). c,d, Results from QMC simulations. c, The upper bound NBG/N on the BG fraction is shown as a function of Δ for s = 11ER. The error bars show the standard error in the mean for the QMC statistical noise. The insets are three-dimensional contour plots of the highest (blue) and second highest (green) occupation eigenfunction of the single-particle density matrix for Δ = 0.05ER (upper left) and Δ = 1ER (lower right). The BG phase has multiple localized modes, only one of which is shown here. d, Density slices through the trap centre are shown for Δ = 0.1 (i), Δ = 0.6 (ii), and Δ = 1ER (iii). The blue (green) regions are the SF (BG) domains, and the colour bar shows the average number of particles on each site.

To quantitatively characterize the amount of excitation present after the quench, we measure

where is the measured OD at the pixel indexed by i and j, and f is a smooth fitting function (see Supplementary Methods). This method was previously used to observe the quantum Kibble–Zurek (KZ) effect by measuring excitations generated through a quench between Mott insulator (MI) and SF states in a ‘clean’ lattice17. Data for s = 11ER and Δ0 ≈ 0–1ER are shown in Fig. 2a. It is apparent that excitations are not generated by the quench until a threshold disorder strength is crossed, above which increases approximately linearly with Δ0. Similar threshold behaviour is observed for all s we sample in this work.

One way to account for such behaviour is by excitations created through SF puddles merging as the BG–SF boundary is crossed18. As spatially distinct SF puddles combine, the random phases associated with each island naturally lead to vortices and other excitations. This process and the threshold behaviour evident in Fig. 2a is generally described by the quantum KZ effect in the context of QPTs (ref. 19). In the KZ scenario, an adiabatic transition from a phase-disordered (for example, BG) to an ordered (for example, SF) state is impossible because of diverging characteristic scales of length and time in the vicinity of the transition. Dynamically traversing a continuous QPT by tuning (or quenching) a Hamiltonian parameter leads to the formation of excitations that persist even after the transition is crossed. The quantum KZ effect has been used to describe this phenomenon, and it has been observed for the MI–SF transition in a ‘clean’ optical lattice, where sensitivity to the equilibrium phase boundary and power-law scaling were observed17, and the scaling of the coherence length after the quench with the quench rate was measured20. Our knowledge of how KZ physics is modified by disorder and glassy phases is limited to simulations and theory of certain one-dimensional spin models3,21,22. The prime example of a quantum quench in a disordered system before this work was in an Ising magnet23, where a smaller residual energy for quantum annealing compared with thermal annealing was observed by means of magnetic susceptibility measurements.

To connect the observed threshold disorder with the SF–BG transition, we carry out exact quantum Monte Carlo (QMC) simulations of the equilibrium system using the same trap and lattice parameters, atom number, and speckle disorder as in the experiment (see Supplementary Methods). For trap-free geometries in the thermodynamic limit, the BG is characterized by a vanishing superfluid order parameter and non-zero compressibility24. In contrast, the trapped system we consider exhibits domains corresponding to SF and BG phases that we distinguish using the spatial extent of the condensate. The condensate is identified as the macroscopic occupation of a single-particle eigenstate that we can obtain from the single-particle density matrix (refs 25,26).

For clean systems (that is, Δ = 0) with U/t < 29.34 ± 0.02 that are below Tc, a single condensate extends throughout the system that coincides with the local superfluid density order parameter (see Supplementary Methods). As Δ is increased, this behaviour changes, and the extent of the macroscopic condensate shrinks, leaving behind regions devoid of coherence. Because the SF–BG transition is of the continuous type, phase coexistence is forbidden, and we identify these regions as BG phase. To illustrate how the BG phase emerges in the gas, we show the two highest occupation eigenfunctions of ρ1 for s = 11ER and Δ = 0.05ER and 1ER in Fig. 2c. At low Δ, all single-particle states are spatially overlapped with the SF domain, and the second highest occupied state results from interaction-induced quantum depletion. For sufficiently high Δ, however, this extended state is replaced by a spatially localized mode that corresponds to a non-macroscopic and locally coherent superfluid puddle, characteristic of the BG phase.

To compare with the measurements, we compute the BG fraction NBG/N as the fraction of atoms in regions without a macroscopic condensate present. This estimate is an upper bound at non-zero temperature because of thermal excitation, which we find is small in the regime we study (see Supplementary Methods). As shown in Fig. 2d, the BG as defined by this criterion emerges at the edge of the gas and grows in extent and number as Δ is increased. The typical behaviour for NBG/N at s = 11ER as Δ is varied (shown in Fig. 2c) mirrors the amount of excitation created by the quench in the experiment: NBG/N is only non-zero above a threshold disorder, above which it increases approximately linearly with Δ.

We construct a SF–BG phase diagram (Fig. 3) by estimating the threshold disorder Δth for excitations to appear in the experiment and for BG to appear in QMC simulations using a piecewise-linear fit to data such as those shown in Fig. 2. The fitting function assumes constant behaviour for disorder strengths less than Δth, and linearly increasing behaviour characterized by the free parameters Δth and a slope for disorder strengths greater than Δth. Several important features of the phase diagram are evident. The threshold disorder Δth is weakly dependent on U/t, and the QMC and experimental results agree within the 25% systematic uncertainty in Δ0; there are additional systematic and statistical uncertainties arising from finite temperature and disorder averaging (see Supplementary Methods). This agreement demonstrates that the quench dynamics and production of excitations in this strongly disordered system are sensitive to the ground-state, equilibrium phases (which are not significantly distorted by non-zero temperature), and that the QMC simulations accurately describe the physical experiment. Moreover, it implies that the SF puddles present in the QMC simulations (Fig. 2c) exist in the experiment. The observed threshold behaviour cannot be explained by mean field theory, which predicts that a BG appears for infinitesimal disorder27. The decrease in Δth at higher s (that is, larger U/t), which cannot be accounted for by general classical percolation mechanisms, indicates that stronger interactions weaken the SF against localization in this regime. Finally, the observed Δth is lower than that required to convert the entire gas to a localized BG phase, as determined through transport measurements13. This behaviour is expected, because the entire gas must be converted to a BG phase in order to halt transport.

Figure 3: Phase boundary between SF and BG regimes.
figure 3

Every point is the result from a piecewise-linear fit to data at fixed s (such as those shown in Fig. 2a and c), and the error bars show the fit uncertainty. The solid black circles are the experimentally determined values of Δth from quench measurements, and the open red circles are the QMC simulation results.

We explore the dynamical timescale of the SF–BG transition by varying the quench time τQ at fixed s and Δ0. In the KZ scenario for clean systems, the amount of excitation and heat produced during a quantum quench typically exhibit power-law dependences on the quench time19. The knowledge of how this changes in disordered systems is restricted to theory and simulations of one-dimensional spin chains, which possess an inverse logarithmic dependence of the residual energy (that is, the energy generated by the quench) on the quench time21,22. Data for nearly two orders of magnitude in τQ/(/t), where /t is the tunnelling time, are shown in Fig. 4 at s = 10ER for the BG regime at Δ0 = 1ER. To avoid complications from decay of excitations during the quench, we measure the temperature of the gas after allowing rethermalization in the trap for 150 ms (see Supplementary Methods). We show the residual energy as the change in the thermal energy δε = εε0, where ε and ε0 are the thermal energy per particle inferred from the measured temperature with and without the quench.

Figure 4: Dependence of the residual energy on the quench time.
figure 4

The change in the thermal energy per particle δε is measured for varied quench time τQ, which is normalized to the (fixed) tunnelling time /t. Data are shown for Δ0 = 1ER, which is greater than Δth and corresponds to the BG regime. The error bars represent the standard error in the mean for the 8–15 measurements averaged for each point. The solid line is to a power law, and the dotted line is a fit to an inverse logarithm. The reduced χ2 values are 0.9 and 0.8 for the inverse logarithm and power law, respectively.

The extremely weak dependence of δε on τQ is characteristic of the BG regime for all s sampled in this work. The residual energy is nearly fixed for quenches spanning approximately six to sixty times the tunnelling time. We do not observe significant changes in this behaviour with different s. The range of Δ0 we can sample is limited to approximately 1ER at s = 10ER—and less at higher s because this measurement requires the heat generated by the quench to be greater than that from the lattice light, and for the gas after the quench to remain condensed. Constraining the scaling with τQ is challenging given the observed weak dependence and our inability to explore longer quench times because scattering of the lattice light heats the gas above Tc. The fits to an inverse logarithm proportional to 1/log(τQt/)1.0±0.4 and a power law (τQt/)0.4±0.1 shown in Fig. 4 are equally consistent with the data.

Theoretical predictions for how the residual energy changes with quench time in this system are unavailable, as exact numerical simulation of a quench is intractable for experimentally relevant numbers of particles in two and three dimensions. Approximate simulations of other dynamics have been carried out, including exploration of how modulating the t leads to excitations28. A clearer understanding of precisely how the residual energy in disordered quantum systems scales with quench time is critical to applications such as quantum annealing, which can be used for many NP-hard optimization problems that map onto disordered quantum systems5,7,29.