Multi-coil (MC) shim arrays [1] provide an alternative to spherical harmonic (SH) shim coils for compensating static and dynamic high-order background field (B₀ ) variations caused by local susceptibility gradients in the brain. In contrast to SH coils, MC shim arrays can be rapidly switched with minimal eddy currents using low-cost op-amp drivers. However, as originally realized, MC shim arrays compete with RF receive arrays for space close to the head, where both systems perform with highest efficiency. Integrated RF-shim arrays [2-3] overcome this limitation by using the same physical wire loop for both RF and DC shim currents, without compromising the performance of either system. Prototype 32ch RF-shim arrays (Fig. 1) at 3T show greatly improved shimming versus standard 2nd-order shims, reducing geometric distortion in EPI (Fig. 2) and thus the need for parallel imaging approaches (e.g. GRAPPA [4]) while still permitting access to these methods when needed. In this work, we revisit the question of how best to combine RF and shim loop geometries on a close-fitting helmet. While optimized shim loop designs have been previously explored for cylindrical geometries [5,6], and a method exists for dividing large RF body loops into multiple shim loops [7], to date no one has shown how to improve the efficiency of integrated RF-shim helmet arrays while preserving RF coverage, sensitivity, and neighbor-overlap decoupling [8]. In this work, we propose a “hybrid” RF-shim array in which shim-only loops are added over the face for targeted shimming of the frontal lobes. We also investigate how to optimize the performance of “compressed arrays” in which only an optimized subset of RF coils are used for shimming, an important practical consideration for reducing hardware complexity.Shim performance is simulated using gradient-echo B₀ field maps acquired on healthy subjects at 3T after 2nd-order shims have been applied over the whole brain (ΔTE=2.54ms, 62 slices, 2mm slice thickness, 2.4mm in-plane, FOV=24x24cm). Biot-Savart code [9] is used to calculate the B₀ offset field generate by loops in arrays patterned on a close-fitting helmet (8-to-128 elements) along with the original 48ch cylindrical shim-only array [1] for comparison. The Matlab function ‘fmincon’ is used to find currents in each loop that minimize the least squares B₀ deviation, $$$\Sigma_{voxels}\big|B_0 -B_0^{COIL}\big|^2$$$, subject to constraints on the max. current per loop (2.5A) and over the whole array (60A). (A.) FACE LOOPS: Starting with a “standard" 32ch array, eight shim-only loops are added over the face and the position, diameter, and number of turns is varied to find an optimal solution. Shims are computed on both a global (whole-brain) and slice-optimized basis. (B.) ARRAY COMPRESSION: The Matlab genetic algorithm 'ga' is used to choose globally optimal subarrays of various sizes on the 40ch array designed in part (A.). The algorithm represents each candidate subarray as an binary vector and then calls 'fmincon' to solve for the least squares optimal shim currents for each candidate design. The optimization is performed over six different subject ΔB₀ maps to control for inter-subject variability. The objective function penalizes σ_B0 over the whole brain and is halted when the design stops “evolving” between iterations.FACE LOOPS: The optimal face loop design uses 65mm dia. coils arranged symmetrically (Fig. 3). The 8 added loops reduce $$$σ_{B0}^{GLOBAL}$$$ by 16% and $$$σ_{B0}^{SLICE-OPT}$$$ by 8%, defined as the standard deviation of ΔB₀ over the whole brain for the two simulated cases. In frontal lobes the impact is more dramatic, rivaling the performance of a 64ch shim array. Four wire turns were considered sufficient, with additional turns yielding only marginal improvements while introducing unwanted copper near the RF coils. ARRAY COMPRESSION: The placement of coils in the optimal subarrays shows a high degree of right-left symmetry and heavily favors the face loops (Fig. 4). As the array grows, each additional element brings diminishing returns for reducing σ_B0. Using half of the available coils, a 20ch coil reduces σ_B0 from 67 Hz to 46 Hz as compared with 44 Hz for the 40ch case, an improvement of only 3%. We show that augmenting a 32ch integrated RF-shim array with few added shim loops over the face offers markedly improved performance without introducing enough copper to interfere with RF receive performance. This creates a “hybrid” array that combines the benefits of integrated RF-shim arrays with multi-coil shim-only arrays. We further show an algorithm for “channel compression” that selects optimal subsets of shim loops from a given array of available coils, thereby reducing hardware complexity. Software for these calculations are available at http://rflab.martinos.org